PLUR1BUS — Memory plugin for OpenClaw

PLUR1BUS turns OpenClaw into an agent with long-term memory: a per-agent isolated LanceDB store as the source of truth, a mirrored Obsidian vault as a human-readable view, and a small set of background jobs that classify, consolidate, and (when warranted) notify.

PLUR1BUS 7.12.8 — verified on OpenClaw 2026.8.x, 2026.9.1 and 2026.9.2

Current source version: 7.12.8. PLUR1BUS 7.12.8 supports OpenClaw 2026.8.1 as its primary host target and is additionally verified against OpenClaw 2026.9.1; the declared compatibility floor is openclaw@2026.8.1 and plugin API >=2026.8.1. The package is built and tested against the immutable build baseline openclaw@2026.8.2; see the compatibility contract for the full runtime matrix and evidence. The upstream source base is the immutable official tag v7.4.10, commit c0a8a4c28ff1cb9c632e185f21f4502d67d1b605.

Web interface

PLUR1BUS registers its own "PLUR1BUS" tab in OpenClaw's Control UI (/plugins/memory-lancedb-namespaced/control) — memory health, the workspace matrix, provider status, and migration progress, all in the browser. It rides on the OpenClaw Gateway's own port and authentication (no separate port, no separate login); reach it through however you already reach your Gateway (loopback, or a Tailscale/VPN front end), then open that tab.

What it does

By default, each agent gets its own LanceDB store under {baseDbPath}/{agentId}/ and a matching Obsidian vault folder for browsing. An explicit named-namespace configuration can read the same validated agent from multiple storage namespaces while keeping one active writer. The plugin captures conversation-derived memory cards automatically, runs a daily consolidator and a critical-push classifier as cron-driven background jobs, and exposes a small set of Telegram commands so the user can inspect, edit, or toggle behaviour without leaving the chat.

New in v7.12.38 — the persona voice learns daily, from evidence

Persona evolution used to run once a week, hand the model only the current profile and ask for "one small change"; the reply outcomes were merely a gate. Now it runs daily (04:15, staggered per agent, personaVoice.cron/timezone to override) with two brakes kept in .adaptive-learning/persona-evolution-state.json: at least minDaysBetween days (default 2) since the last change, and at least minOutcomes (default 10) new real reply outcomes since then, more than half of them positive. Heartbeat turns no longer count. The model sees the protected seed, the numbered learned lines, and up to six positive and six negative excerpts (question, answer, the user's reaction) and answers with ADD: - …, REPLACE <n>: - … or NONE. The managed block may hold 24 lines (personaVoice.maxBullets, was 12 including the seed) and the injected directive cap follows from that (lines × 130 + 80, default 3200 characters). Existing weekly jobs are migrated to the daily slot on the next setup run.

Twenty-five releases in three days, all live-tested on a three-agent installation. The changelog has the details per version; the highlights:

• Prompt hook no longer waits on the database. OpenClaw runs a plugin's before_prompt_build handlers one after another, each with its own 15 s timeout. Reply-outcome tracking ran first and waited for up to twelve sequential LanceDB updates, so recall started only after that timeout. The DB work now runs in a serial per-agent queue after the turn's recall (7.12.30). A new recall prelude log line reports where the pre-recall time goes; the query embedding runs under a budget (neo.recall.global.embedTimeoutMs).
• Hook identity works with OpenClaw 2026.9 session entries (7.12.30, 7.12.32, 7.12.33). Session entries moved channel, account and target under entry.delivery; the resolver reads both the old and the new shape. Voice notes and photos without a caption now produce a dispatch ticket, and the fallback warning names the exact step and reason.
• Neo store on a vector sidecar (7.12.26). Embeddings live as raw float32 in vectors.<gen>.f32 with byte offsets in vector-index.json instead of JSON number lists; a 200 MB candidate journal shrank to about 13 MB plus a 38 MB sidecar. Lossless; inline vectors keep being read. scripts/migrate-neo-vectors-sidecar.mjs migrates existing stores.
• Global candidate search (7.12.27, 7.12.28). Recall searches every candidate of the workspace by cosine on the sidecar, with a recency discount and de-duplication against injected LanceDB memories (neo.recall.global). A metadata index (candidate-index.jsonl) keeps the horizon independent of the journal cap and makes the search cost 60–140 ms after the first read.
• Nightly LanceDB compaction (7.12.31, 7.12.33). Every row update creates a fragment and a version; tables reached hundreds of one-row fragments and thousands of versions, which made vector search and updates ten times slower. consolidate-daily now runs optimize() per agent (dailyConsolidation.lancedbOptimize, versions older than 24 h are pruned).
• Daily consolidation repaired (7.12.24, 7.12.29). Graph edges whose memories are gone are pruned; the memory-dynamics decay scans only the agent's own partition instead of failing on the first foreign-scope row; ACL denials carry a code, a reason and the row id.
• Emotion tier 3 moved to a cron (7.12.22, 7.12.23). Capture runs tiers 1/2 only; the hourly emotion-refine job refines rows the fast tiers were unsure about, and always rows with importance ≥ 0.9.
• Cold-start and worker fixes (7.12.20, 7.12.21, 7.12.24, 7.12.25). Stable turn identity for capture, a persisted cap-check watermark for the journals (first append after a restart dropped from 17 s to under 1 s), worker warm-up per plugin instance, worker timings in the capture log.
• Operator path for chat commands (7.12.18, 7.12.19, 7.12.24). openclaw plur1bus-command --agent <id> --session <key> [--locale xx] "/…" runs any chat command as operator against the same Neo store the hooks use.
• Reminder nudge loop closed (7.12.24): a presented reminder can no longer be captured as a new reminder.

New in v7.12.8 — the KNOWLEDGE.md queue drops entries it can never promote

• knowledge_update refuses to promote an invalidated memory into canonical KNOWLEDGE.md, but afterwards removed only the keys it had actually integrated. Anything unpromotable — invalidated (a soft delete sets both status="deleted" and epistemicStatus="invalidated") or gone from the table — stayed queued forever in .adaptive-learning/knowledge-pending.json and kept inflating pendingCount, so the maintenance nudge told the agent that N insights were waiting when some of them no longer existed. The queue now drops those entries on the next run, before the "nothing to integrate" exit. Pruning is restricted to the ids the query actually asked for, so a capped query or a failed fetch never drops pending work.

New in v7.12.7 — MEMORY.md and USER.md reach the session again, embedding calls time out

• OpenClaw asks the memory-slot owner to classify MEMORY.md and USER.md before injecting them (classifyWorkspaceMemoryPaths). PLUR1BUS did not implement it, the host answered unsupported, logged "excluding automatic memory context: selected memory runtime does not support provenance classification" and left both files out of every session — the agent only saw its curated memory when it read the file by hand. The runtime now classifies workspace memory paths with the same rules as memory-core (curated roots and memory/**.md → agent, dreams → system, anything outside the workspace or non-memory → untrusted).
• Remote embedding requests carry a per-request timeout (embedding.requestTimeoutMs, default 15 s) and no SDK-side retries. The OpenAI SDK default of ten minutes let a stalled request hold a recall until the 20-second worker timeout (started=yes elapsedMs=0 in the log).

New in v7.12.6 — recall returns memories again (BigInt expiry rows were treated as expired)

• LanceDB returns Int64 columns (expiresAt, validFrom, validUntil, updatedAt, remindAt, …) as JavaScript BigInt. The recall pipeline's lifecycle gate accepted only numbers, so a BigInt 0n expiry counted as "expired" and every candidate was rejected before scoring, ACL and reranking. Auto-recall injected zero memories on every turn and memory_recall reported "no results" while the rows were in the table with scores around 0.5. projectRecallEntry now coerces all Int64-backed fields to safe numbers and isRecallEntryLive accepts BigInt input. Direct MemoryDB.search was unaffected, which is why manual checks kept passing.

New in v7.12.5 — umlauts survive query refinement, recall timeouts name the slot state

• The recall query refiner decomposed text with NFKD and then stripped the combining marks, so every German query with ä, ö or ü was split apart (Gespräch became Gespra ch, läuft became la uft) and umlaut stopwords such as über or für never matched. Combining marks are kept and the text is recomposed to NFC.
• When a recall times out before it ever received a worker slot, the warning now says so (started=no), reports the wait time, the active slot count and maxConcurrentRecall, and the job is dropped from the queue instead of running uselessly later. This is the signal to watch when runtime.maxConcurrentRecall is too low for the host.

New in v7.12.4 — the transcript is undated, semantic-discover script imports homedir

• The <temporal-context> block now states that transcript messages carry no timestamps and compaction summaries no dates, and that no time or date may be asserted for earlier conversation content unless it comes from the block, a memory record's created-at/age attribute or a tool result. Agents had been filling that gap with invented times (#133).
• scripts/run-semantic-discover-once.mjs used homedir() without importing it from node:os and crashed unless PLUR1BUS_VAULT_PATH and PLUR1BUS_DB_BASE were set. The import is in place.

New in v7.12.3 — KNOWLEDGE.md promotes again, dreams reach the agent's workspace

• schicht15.maxPromotionsPerRun compared the lifetime number of promoted cards against the limit, so a workspace that had ever promoted that many was blocked for good (a live install sat at "7/3" since June with KNOWLEDGE.md frozen). The limit now counts a 24-hour window; promotions recorded before timestamps existed do not count, so a stuck workspace unblocks on the next run. Dedup of already promoted cards is unchanged.
• The agent's own REM partition wrote its DREAMS.md diary, vault note, mood and soul context into its Neo store directory instead of the agent workspace, so the host's Dreams page never showed PLUR1BUS dreams and the narrative ran without mood. The agent's own private partition now writes to its workspace; other agents' partitions and user pools stay in their store directory. Light dreams accepted only the scope spelling "agent" for the diary while ACL bindings say "agent-private"; both are accepted now.

New in v7.12.2 — critical pushes show what they are about

• A critical push for a health or finance card used to read "The content is hidden for privacy reasons", leaving the owner to accept or reject a card they could not see. The push goes to the owner's direct chat only and quotes the owner's own statement, so the sanitized 160-character preview is now shown for every type except credentials, which stay hidden. Operators who want the old behaviour list the types under criticalPush.hideTypes (for example ["gesundheit", "geld_konto"]).

New in v7.12.1 — macOS portability for the scoped embedding owner and the test suite

• The scoped embedding owner always bound a Linux abstract Unix socket, which macOS rejects with listen EINVAL before the owner service starts. Linux keeps its exact address; other Unix platforms elect the owner through a deterministic exclusive loopback TCP claim that carries no data, while the data plane stays the private filesystem socket with its 256-bit token. An oversized macOS socket path (103 bytes) is diagnosed before any mutation.
• Tests compare workspace and generation paths canonically (/var is /private/var on macOS), skip the four shared-memory integration paths on hosts without stable fd-backed directory routing with an explicit reason, and assert that such hosts reject shared writes without creating a shared root. Two focused macOS CI jobs run next to the unchanged Linux suite.
• Also verified on OpenClaw 2026.9.2 (full suite against the 9.2 host: same three deliberate baseline failures as on 9.1, nothing else).

New in v7.12.0 — Jina v5 Text Nano is the proposal for new installs

• After the lab test (see the comparison below) the provider wizard and the shell installer offer jinaai/jina-embeddings-v5-text-nano-retrieval as option 1 and default; OpenAI is option 2, E5 the keyless fallback, Jina v3 stays selectable for existing installs. Non-interactive and dry runs need PLUR1BUS_ACCEPT_NONCOMMERCIAL_LICENSE=1 to accept CC BY-NC 4.0 and fall back to E5 otherwise, rather than failing or acknowledging a license silently.
• Existing installs are not touched. The dashboard's Embedding Dimension Planner now recommends the migration to Nano when another embedding model is active: prepare the target in Model Preparation, then run the re-embedding migration (dry run, copy, separate switch); the old generation stays for rollback.

New in v7.11.1 — long cards no longer drive the local Jina models into the OOM killer

• Both pinned Jina models accept 8,192 tokens, and the ONNX runtime keeps the attention working set of the longest text in a batch, times the batch size, without ever returning it. In the 7.11.0 lab run the v3 process grew to 41 GB and was killed at card 808 of 1,031; one batch of the eight longest cards (up to 15,000 characters) took nano to +30 GB and 117 s. Both Jina runtimes now cap every text at embedding.local.maxTokens, default 512 (32 to 8,192). The same batch then costs +0.6 GB (nano) or +1.1 GB (v3) and 5 or 22 s. Memory cards are summaries; raise the cap deliberately if long cards must be embedded in full. The cap is part of the shared model pool identity.

New in v7.11.0 — Jina v5 Text Nano as a local embedding option

• The published Q8 ONNX export of jinaai/jina-embeddings-v5-text-nano-retrieval is pinned (five artifacts, about 265 MB, SHA-256 checked). A jina-v5 runtime branch validates the EuroBERT config, takes the graph's own sentence_embedding (normalized last-token pooling) or pools over the attention mask itself, truncates to 32 to 768 Matryoshka dimensions, and distinguishes queries from documents by the published Query: and Document: prefixes, refusing any other prefix. Targets jina-v5-nano-* appear in the schema, the dimension planner, the dashboard switch (same re-embedding migration), wizard option 4 and installer choice jina5, each behind the CC BY-NC 4.0 acknowledgement. An option, not a proposal: the default stays where it is until the PLUR1BUS lab test against v3 is in, and existing configurations are untouched.

New in v7.10.0 — BGE is the recommended local reranker

• The provider wizard and the shell installer now offer the local BGE reranker as option 1 and the default; Cohere is option 2, Jina v2 stays under the advanced options. The choice is selected by key, so the order can change without shifting the branches. The reason is documented under Choosing embedding and reranking models: on the multilingual benchmarks that matter for memory cards BGE is ahead or level, and it needs no license consent. Nothing changes for an existing configuration; the runtime default was BGE already.

New in v7.9.2 — tombstoned cards are no longer classified or pushed

• A card removed with memory_forget keeps type = "memory" while its status becomes deleted. The classify-recent cron took such rows for fresh candidates, typed them and pushed them as critical, while the review path (which only reads active rows) never knew the references. Candidate selection and the classifier now skip every row whose status is not active; the run result reports them as skippedInactive.

New in v7.9.1 — the quoted reply is caught on the chat dispatch hook

• 7.9.0 listened on before_agent_reply only, which the host does not fire for a Telegram chat message; the message went to the agent. The handler now also listens on before_dispatch, where the host hands over the quoted text.

New in v7.9.0 — accept or reject critical pushes in bulk, also by quoting

• /plur1bus critical accept all, reject all, or several references in one command handle every pending review of the authorized scope; unknown or ambiguous references are reported, the rest is applied.
• Quote a push in the chat and write "bitte alle akzeptieren", "alle ablehnen", "accept all" (or, for a single quoted card, "akzeptieren" / "nicht hervorheben"): PLUR1BUS answers before the agent, taking the references from the quoted push and the decision from the reply, with the same authorization as the command. Only a quote carrying the push's fixed headline counts; anything unclear goes to the agent.

New in v7.8.11 — shared user partitions get names too

• "Cards by user" now labels each pool after the agent, channel and account whose allowed direct-message user owns it (main.telegram.default, numbered when an account admits several users). The user id itself never appears on the page.

New in v7.8.10 — shared partitions get names and an explanation

• "Cards by workspace" showed pool hashes and zeros. The alias snapshot now names each shared partition (main, and main.dir for the pre-alias directory identity of the same workspace), and an all-zero list says that the partitions exist but nothing has been shared into them yet. Paths stay off the page.

New in v7.8.9 — the Obsidian operator CLI works

• openclaw plur1bus-obsidian detect|use|create|confirm --session <key> called the host's RPC helper as (method, params); its signature is (method, opts, params, extra), so every method saw an empty request and nothing was printed. The CLI now passes the parameters in the right slot, with the method's scope, and writes the result as JSON.

New in v7.8.8 — scheduled discovery carries the right capability

• 7.8.6 handed the discovery cron the bridge service's policy, which only carries vault_write; the discovery also needs semantic_index_write, granted by the semantic-discovery confirm plan with a confirmed action. The scheduled run now counts as the operator's standing confirmation; the vault receipt, mode: apply and allowWrite stay required.

New in v7.8.7 — the vault confirmation can be driven from the CLI

• plur1bus-obsidian use/confirm --session <key> and the gateway methods ended with identity_binding_required: the operator-side session resolver produced only agent and workspace, no user or conversation identity. A direct-chat session key now resolves to the same identity-bound context that conversation's chat commands get, so the receipt binds to that conversation's user. Other session kinds keep the plain agent context.

New in v7.8.6 — scheduled semantic-link discovery can write

• The discover-semantic-links cron called the discoverer without a mutation policy, which the policy layer reads as blocked, so scheduled discovery never wrote a link. The handler now builds the same receipt-bound policy per workspace that the bridge service uses. Without a vault receipt it stays blocked by design.
• The vault confirmation flow (plur1bus-obsidian use/confirm, gateway methods plur1bus.obsidian.prepare/confirm) failed with "Invalid agent ID": the handlers read the parameters off the host's method context object and did not accept the CLI's session name. Both are handled now, and the agent id falls back to the key's agent:<id>: prefix.

New in v7.8.5 — the Obsidian card sees every agent's vault

• Without a configured obsidianBridge.workspaces list the target probe only looked at <OpenClaw home>/workspace; the vaults in workspace-<agent> stayed invisible. Every workspace* directory is probed now.
• With requireVaultPathConfirmation: false the bridge acts without a receipt, so a configured target now reads "ready" instead of "not yet confirmed", and the card says how many targets are configured.
• The bridge only knows vaults through obsidianBridge.workspaces (entries with id, agentId, path) or vaultPath. The reviews work without it through the host's agent workspace; semantic-link discovery and the target card do not.

New in v7.8.4 — the dashboard buttons work inside the host tab

• OpenClaw embeds the plugin tab in an iframe sandboxed with allow-scripts only, so the browser silently blocked every native form submission there; the 7.6.0 switches and the 7.8.0 Compact button only worked in a standalone browser tab. A writable page now carries one nonce-bound script that posts the form with fetch and reloads itself; the result shows up as the usual banner. Read-only pages still carry no script. The script intercepts the click on the submit button: a sandboxed frame refuses the submission before it ever fires the submit event (7.8.1 listened there and never ran). The action travels as a GET with the single-use token in the query, because the host authenticates the frame's cookie for GET only and answers a POST from the opaque origin with 401 (7.8.2 hit that). 7.8.4 routes the Compact button through the db-adapter, which owns optimizeTable; 7.8.3 asked the pool's raw store objects and failed.

New in v7.8.0 — a Compact button per partition, and an installer that reaches its plan

• Compact behind every LanceDB row. With controlUi.writeActions: "all", each private partition under "Cards by agent" gets a button that runs LanceDB's fragment compaction (table.optimize()) for that partition in the background, one at a time. The adapter had the primitive since August; nothing ever called it. The row shows progress and the result.
• The feature-cron installer no longer times out its own probe. On large installs openclaw plur1bus-feature-cron --help takes longer than the old five-second budget, so every run took the fail-closed branch and safety-disabled the direct jobs. Budget is 30 s now.
• Ownership rules fixed: a disabled owned job no longer counts as present for non-delivery features (it is re-enabled), operator-named consolidations that run the shipped command are not duplicated, the singleton collector is recognised on any agent, and main is the default agent when the host flags none.
• Memory Health no longer reads "degraded" because of _neo. Reserved store directories are skipped by the health scan.

New in v7.7.0 — dreams on OpenClaw's own Dreams page

• PLUR1BUS dreams show up in Settings → Memory → Dreams. Light and REM narratives are written into the agent's DREAMS.md, in the host's entry shape inside its managed diary block. Only the agent's private partition writes there; shared partitions never do. dreaming.narrative.diary turns it off. To give the diary one author, switch the host's own managed dreaming off. The host reads that switch from the memory slot owner's entry, so it is plugins.entries.memory-lancedb-namespaced.config.dreaming.enabled: false, not memory-core's flag.
• The Memory page knows the engine. PLUR1BUS now registers the memory runtime the host expects from the slot owner, so Overview and Scene report PLUR1BUS' provider, model, embedding state and card count instead of "memory plugin unavailable". Host-side search runs through PLUR1BUS' recall pipeline on the agent's private partition only.
• Not reachable from a plugin: the counters under Advanced and the phase chips, which read memory-core's internal short-term store and its managed cron.

New in v7.6.0 — switch providers from the dashboard

The operator tab can now change two things, if you let it. controlUi.writeActions is off by default, so nothing changes for an existing install.

• reranker makes the reranking choice switchable from the page: local BGE, local JinaAI (both keyless), Cohere when a key is configured, or off. That is a runtime choice with no data migration behind it.
• all additionally exposes the embedding target and the re-embedding migration: dry run, copy, and a separate switch. The confirmation token that binds those steps stays inside the Gateway and never appears in the browser.

Every change carries a single-use form token from the page render it was clicked on, because the tab cookie the host mints is SameSite=None. While writeActions is off the page renders no form, forbids form targets in its Content-Security-Policy, and refuses POST outright.

Also fixed: the reranking card used to name jina as a reranker.provider value. The schema rejects it. Jina reranking is a local model, selected through reranker.local.model.

New in v7.5.6 — verified against OpenClaw 2026.9.1

• No code change was needed for the 2026.9.1 host. PLUR1BUS was loaded through that release's own plugin loader and the full suite was run against it. The compatibility contract now names 2026.9.1 stable and drops the 2026.9.1-beta.1 wording, which described a beta nobody runs.

New in v7.5.5 — readiness agrees with the runtime

• Credential Readiness no longer calls a working embedding "missing". The runtime falls back to OPENAI_API_KEY (and OPENAI_API_KEY_FALLBACK for the fallback embedder) when neither apiKey nor apiKeyEnv is set, which is how a default install carries its key. The table now follows the same rule, names the variable it found, and never reads its value. Keyless providers show not required; an unconfigured optional fallback shows optional.

New in v7.5.4 — the tab opens at once, and looks like the rest of the Control UI

• No more waiting for the health scan. Opening the tab used to run a scan over every partition table (2–21 s depending on gateway load) because the snapshot cache lasted 10 s. The last snapshot is now served immediately and refreshed in the background: once at gateway start, then every 10 minutes. The page says how old the snapshot is.
• Host styling. The page uses the Control UI's own colour, radius and type tokens, dark by default and light when the operating system asks for it. The tab runs in a sandboxed iframe that receives no theme message, so a light/dark mode forced in the host cannot be followed, only the OS setting.

New in v7.5.3 — the tab registers, and a host route is not a defect

• The Control UI tab showed "Plugin panel unavailable". OpenClaw 2026.8.2 exposes the descriptor registration flat on the api object; PLUR1BUS read only the nested path, so no tab was ever registered.
• Host-routed capabilities no longer wear the missing badge. They need no key of their own, so the status says so instead of reading like a fault.

New in v7.5.2 — the readiness table points at the right line

• The config path shown is the one that holds the key. A key in reranker.apiKeyEnv was labelled reranker.apiKey, sending the reader to a line that does not exist in their config.
• host_route is explained. The source had no legend entry, and the missing help text claimed the feature stays off — untrue for exactly the capabilities that report it.

New in v7.5.1 — the operator dashboard tells the truth

• Continuity Engine, Semantic Lens and Query Refinement default to on. The tab states that an absent switch means on, but these three defaulted to off in the manifest, so an untouched install contradicted the sentence printed above the feature cards. An explicit enabled: false still turns them off.
• Credential readiness reads apiKeyEnv too. A working Cohere reranker configured through an environment variable was reported as missing. The five capabilities that fall back to OpenClaw's own model route are now marked as host-routed instead of looking unconfigured.
• A Reranking card. Provider, model and revision of the active reranker, plus how to switch between cohere, jina and the local BGE fallback.

New in v7.5.0 — native integration and verified local models

• No OpenClaw bundle patching. Feature cron commands use OpenClaw's public plugin registration and gateway-runtime dispatcher capabilities. Missing capabilities fail closed; PLUR1BUS never rewrites OpenClaw source, dist, or node_modules files.
• Exact 2026.8.x release contract. Package metadata pins the tested host and SDK identity to 2026.8.1/2026.8.2, while runtime behavior is guarded by feature detection rather than version-string branches. OpenClaw 2026.9.1 stable is additionally verified (see the compatibility contract).
• Pinned local inference. E5 embeddings, the optional multilingual Jina v3 embedding, Jina reranking, and the free BGE fallback use immutable Hugging Face revisions with exact required-file sizes and SHA-256 verification before Transformers.js loads them. Jina v3 embedding is offered separately because its CC BY-NC 4.0 license does not permit commercial use without other terms.
• Activation-owned preparation. Selecting a local-model preparation profile stages no work during plugin discovery. The download begins only after OpenClaw activates the PLUR1BUS service, and shutdown or replacement drains it through the public service/runtime lifecycle.
• Lifecycle-owned Obsidian watcher. OpenClaw starts the Bridge only after registry activation and stops it before replacement. Periodic scans queue inbound review candidates without starting a Memory write; an authorized explicit apply remains required for durable inbound mutation.
• One local model across scoped registries. OpenClaw's request-scoped provider registries delegate local embedding inference to the activated PLUR1BUS owner over a private, authenticated Unix socket below the PLUR1BUS state directory. It is never published as a host port, binds exact model and dimension identity, and closes before the activation-owned model.
• Deterministic regression timing. CPU-bound performance contracts measure process CPU time; repair diagnostics drain naturally before process exit.

New in v7.4.0 — evidence the agent can stand behind

• An epistemic status every write earns. New user captures are recorded as observed, every other new write as explicit untrusted; nothing invents trusted any more. The skill miner clusters observed | corroborated | trusted plus valid pre-cutoff legacy rows and no longer applies a 30-day lookback, so an existing install keeps mining its history instead of reporting scanned: 0. The cutoff marker is written once, at the first upgrade, before the first write.
• Skill approval survives a crash. SKILL.md is written first (tmp + fsync + rename), then the evidence transitions; a partial failure stays activation_partial and can be re-applied idempotently.
• Forgotten stays forgotten. Every reachable card re-insert — store, content update, updateCard, compaction, auto-capture, light-dream rewrite — checks the tombstone registry before table.add. Same-text replay by the user is still allowed.
• A global inject budget (recall.globalInjectMaxChars, default 17000) trims memories before time and reminder context, so a large recall can no longer crowd the rest of the prompt out.
• Two new curation commands. /plur1bus curation resolve <keep|drop> ends a neo conflict without any hard filter, and /plur1bus curation drop-injected demotes only injected behaviour conflicts after a preview and a nonce — genuine conflicts are never touched.
• Derived records carry visibility. Pattern and dream writers stamp scope; readers filter by requester. Legacy records without a stamp stay own-agent only.
• Legacy compatibility note. The 7.4.0 release could opt out of its then current dispatcher rewrite. Version 7.5.0 removes that rewrite completely.

New in v7.3.x — memory dynamics that actually fire

Condensed summary of 7.3.0–7.3.5; see the changelog for detail.

• Valid-time and trust state (7.3.0) — memories carry temporal validity, and epistemic trust became a first-class field alongside a batch of audit fixes.
• Security and scope hardening (7.3.1) — bound episode-graph endpoints, ownership-partitioned compaction, fail-closed skill scans.
• The classifier stopped rejecting its own default (7.3.2) — fakt was missing from the type enum, so every classification run failed validation.
• GC got a trigger and a policy (7.3.3) — an eighth feature cron at 04:45 plus a configurable maxMemoryCount; before this, garbage collection had no scheduler at all.
• importance = 1.0 works again (7.3.4) — the value is the agent's reserved manual core marker; it was silently ignored because the core score also demanded an emotional intensity the agent cannot set. Core scores are now normalised to the features that actually exist.
• REM dreaming finds patterns again (7.3.5) — similarity was compared on two different scales, so even an identical vector fell below the threshold: zero edges, zero clusters, ever. Feature crons are also staggered per agent now.

New in v7.2.0 — safer OpenClaw updates

• Complete, version-bound deploy verification — repair checks cover all static and literal dynamic runtime imports, bind package and manifest identity to one immutable source snapshot, and roll back the complete deployment if any copy or final validation fails.
• Safe promoted-memory reindex bridge — the replacement maintenance CLI is configuration-aware, namespace-aware, idempotent, dry-run by default, and compatible with predecessor marker state without forwarding redacted credentials.
• Staggered daily consolidation — exact PLUR1BUS-owned per-agent jobs are scheduled 15 minutes apart, avoiding simultaneous consolidation pressure; custom or look-alike jobs remain untouched.
• Explicit recovery tooling — deploy verification and repair accept an expected release version and report missing source/preflight reasons instead of producing a mixed installation.

New in v7.1.9 — lower background token usage

• Six high-frequency jobs reduced by about 83% — Afterthought and Critical Push now run every three hours instead of every 30 minutes.
• Safe schedule migration — only PLUR1BUS's shipped 30-minute cadence is migrated; custom operator intervals and delivery targets remain untouched.
• No active-memory recursion — active-memory child sessions cannot trigger another PLUR1BUS recall or capture pass.
• Agent-scoped semantic discovery — scheduled discovery processes only the triggering agent's configured Obsidian workspaces.

New in v7.1.8 — reliable feature-cron setup

• Slow healthy gateways no longer miss reconciliation — the redacted effective-config snapshot used by feature-cron setup now has a focused 30-second budget; live measurements were approximately 12.5–18 seconds.
• Narrow compatibility fix — the fail-closed result contract, single snapshot, shared CLI defaults, schedules, delivery, model routing, and thinking policy remain unchanged.

New in v7.1.7 — model-free feature cron dispatch

• No outer carrier model for Afterthought or Critical Push — exact internal commands are finalized through OpenClaw's normal delivery path before the carrier agent/model executor.
• Scoped fail-closed recovery — if the host dispatcher is unavailable, only exact PLUR1BUS-owned feature jobs are paused and marked; custom prompts and unrelated jobs are never claimed.
• Schedules and delivery remain intact — existing jobs migrate idempotently while keeping their 30-minute cadence and validated targets.
• Native internal LLM policy — actual Afterthought composition and classification still inherit the target agent's OpenClaw model and thinking policy; PLUR1BUS does not force thinking: off.

New in v7.1.0 — audited ownership, recall, and operations

• Complete high/medium audit remediation (B1–B15) — durable memory writes, timeout settlement, cancellation barriers, diagnostics, installer paths, operational maintenance, Obsidian mutations, and background jobs were hardened and regression-tested.
• Strict memory ownership — private, workspace, and user data use canonical agent/workspace/user request contexts. Sharing is explicit, confirmation-bound, owner-bound, and isolated in separate storage pools; unbound or conflicting rows fail closed.
• Recall and namespace closure — B12 Core and B12-P add secure read-only legacy access, globally bounded multi-namespace recall, adaptive budgets, compression, decision traces, and strict graph/provider authorization.
• OpenClaw owns the LLM choice — PLUR1BUS feature routes inherit the effective target agent model unless a feature has a complete explicit direct-provider override. Hard-coded chat-model defaults were removed.
• Exact LLM result cache — deterministic internal transforms can reuse validated, bounded, agent-scoped results without leaking prompts or credentials.
• Dependency and runtime baseline — patched transitive dependencies and sharp@0.35.3 close the dependency audit; PLUR1BUS now requires Node.js 22.22 or newer.
• Release verification — the release baseline contains 3,260 tests (3,259 passed, 0 failed, 1 skipped) and npm audit reports 0 vulnerabilities.

New in v6.9.10 — Maintenance progress and dedupe hardening

• Candidate status updates survive content dedupe — promote/demote/prune/tombstone updates are append-preserved while ordinary candidate captures remain content-deduped.
• Capped memory dynamics are resumable — retrieval-ledger caps store partial entry progress, and daily decay rotates with a persisted cursor instead of repeatedly touching the first rows.
• LanceDB vector wrappers normalize safely — Arrow-style vector wrappers are converted before update writes to avoid schema failures during consolidation.

New in v6.9.x — Runtime fixes, cron provisioning, and emotional dynamics

• REM-Dream cron provisioning — New installs provision rem-dream when merging.enabled: true is explicitly authored, instead of shipping an enabled handler without a scheduler binding.
• /state command fix — The top-level status command no longer crashes on an out-of-scope ctx reference.
• Emotion config-schema sync — The strict schema now accepts the documented emotional-dynamics keys used by 6.9.x configs.
• Generic temperament defaults — Shipped defaults no longer bake in agent-specific personalities; per-agent temperament belongs in user config.
• Emotional dynamics — Mood persistence, temperament presets, decay modulation, and stronger mood-congruent recall boosts landed in the 6.9.0 line.

New in v6.8.x — Code-review hardening

• i18n sync — 752 missing translation keys added for new OpenClaw channel wizards (IRC, Feishu, NextcloudTalk, Google Chat).
• TypeScript optional dep — typescript is now declared as an optionalDependency so the code-index feature works out of the box without forcing TS on all users.
• Installer fixes — buildInstallLogEvent now correctly passes featureMode instead of hardcoding "preserve"; dry-run vs. remote-target warnings are properly distinguished; dead code removed.
• Installer performance — 9 sequential jq subprocess calls consolidated into batch eval+@sh extracts.
• Neo worker drain — Missing await on drainEmbeddingQueue() caused the unresolved Promise to be serialised as {} in postMessage; callers now receive correct drain results.
• Auto-capture robustness — statSync race condition fixed (file deleted between readdirSync and statSync); addQueryVector null-return guard added.
• ts-source-indexer — O(n) symbols.find() in AST visitor replaced with a Map for O(1) lookup.
• Manifest sync — openclaw.plugin.json version aligned with package.json.
• Security hardening — scope: "user" writes now require an authenticated user identity (user scope is owner-bound) and are filtered in recall/visibility checks.

New in v6.8.0 — Release readiness, code context, and runtime packaging

• Media diarization context — Async diarization merge pipeline, manual speaker mapping, contextual speaker-name proposals, and no biometric enrollment.
• Emotional-state injector packaged — Tracked .openclaw/extensions/emotional-state-injector/ files are included in the npm tarball; runtime activation still requires the OpenClaw plugin entry/allow config and a gateway restart.
• Performance follow-up — Legacy auto-capture duplicate checks are batched, duplicate lookup can use ANN multi-query search when available, JSON hot-path writes are queued asynchronously, and high-cost prompt work is narrowed.
• Optional code index — Local JS/TS index generation writes .plur1bus/code-index.json and can render bounded <code-context> query output.

Experimental code index

PLUR1BUS can build a local JS/TS code index without CocoIndex:

npm run code-index -- /path/to/workspace
npm run code-index -- /path/to/workspace --query "/plur1bus code-index"

The index is written to .plur1bus/code-index.json and contains normalized files, symbols, import/call/register edges, and symbol chunks. --query prints a bounded <code-context> block from the generated index. It uses the TypeScript Compiler API through the optional typescript dependency and keeps the PLUR1BUS schema independent of the parser implementation.

New in v6.7.x — Multi-Namespace, Temporal Continuity & Source Sync

• Multi-Namespace Pool — One validated agent can opt into recall across named LanceDB storage namespaces; this never selects another agent and is not cross-agent sharing.
• Temporal Continuity Context — Auto-injected time-anchor block lets the agent orient itself after gaps or compactions without hallucinating dates.
• Conflict Summary Management — Contradiction detector now emits structured conflict summaries; /plur1bus obsidian conflicts build renders them as Obsidian pages.
• /plur1bus start onboarding — Read-only status and onboarding guidance for feature profiles and vault setup.
• Auto-capture schema sync — scripts/auto-capture-lancedb.mjs gains workspaceKey field; schema migration is backward-compatible and idempotent.
• Internal-turn skip guard — shouldSkipAutoRecallForInternalTurn prevents feedback loops when the gateway injects synthetic cron messages.

New in v6.6.0 — Meta-Cognition

• Self-reflection on recall quality — Precision, Recall, F1 computed from user feedback (/mf +/-/~). Coverage-gap detection finds topics with few or weak memories.
• Threshold-based reflection trigger — Auto-runs when sessionThreshold (default: 50) or intervalDays (default: 7) is reached. Optional LLM-generated natural-language report.
• Persistent state — Reflection state stored in _meta-cognition-state.json per workspace.

New in v6.5.0 — Proactive Nudges

• Embedding-based pattern detection — Clusters similar turns by cosine similarity over embedding centroids.
• Cluster persistence — Clusters survive restarts, stored per workspace/agent.
• Cooldown mechanism — Rate-limited to avoid spam (default: 24h per workspace).
• Configurable thresholds — minClusterSize, similarityThreshold, maxNudgesPerDay.

New in v6.4.0 — Emotion Tier-Config

• Budget-Gate per tier — Tier-1 (regex), Tier-2 (heuristic), Tier-3 (LLM) independently enable/disable.
• Configurable model per tier — An absent model uses the effective OpenClaw agent model; gpt-4o-mini is only an explicit override example.
• Feature-Toggle — Lock emotionTier to a specific tier or use auto for dynamic escalation.
• Graceful degradation — Falls back from Tier-3 to Tier-2 when neither a native OpenClaw route nor a complete direct override is available.

New in v6.3.0 — Explainability & GC

• Explainability (--explain flag for /memory) — Human-readable rationale per result: score breakdown, boost factors, temporal relevance.
• Garbage Collection job — Background cleanup of expired/stale memories with configurable retention policies.
• Feedback Analyzer — Background analysis of /mf feedback for recall-quality improvement.

New in v6.2.0 — Correction-as-Recall

• /correct treated as recall event — After safeUpdate() inserts the corrected card, applyRetrievalReinforcement refreshes lastRetrievedAt, increments retrievalCount, and boosts memoryStrength.
• Null guard — If getById(newId) races or fails, reinforcement is silently skipped; the correction itself is never rolled back.

New in v6.1.4 (Consolidation)

• ACL / Access Control — Agent- and workspace-scoped memory access. searchByTopic, getCard, and recall pipeline filter by ACL. Unauthorized access is logged.
• Feedback loop (/mf) — Thumbs-up/down/neutral feedback on any memory result. Persisted per workspace.
• Temporal reasoning — Queries like "last month", "3 days ago", "Q2 2026" parsed to concrete date ranges before boost/rerank.
• Collaborative memory (/share) — Copy any card into a workspace-shared pool with ACL protection.
• Query refinement — Automatic query rewrite on poor first results, merged and deduplicated.

New in v6.1.2 (Engram — Recall Hardening)

• Recall hardening — maxPromptMemories (default 12), dedup threshold 0.78, acronym recognition, canonicalMaxItems (default 5).
• Typ-based half-life — halfLifeDaysMap: transient (60d), episodic (180d), longContext/project (600d).
• Performance — LRU+TTL embedding cache, semantic recall compression, adaptive recall tiers, graph-index traversal, reinforcement loop.
• Security — SQL-escaping, ACL hardening for destructive commands, path-traversal protection, filter-parser injection resistance.

New in v6 (Base)

• Semantic long-input handling — /memory, /forget, /correct accept any length. >6k chars are semantically compressed; >100k chars prompts for file/vault source.
• Feature activation profiles — On first start proposes a recommended profile (all features active, Obsidian/reviews marked pending_setup). Core memory works immediately; advanced features require explicit confirmation.
• Proposal-only merging — Daily compaction detects duplicates, generates merge proposals in merge-proposals.jsonl, never auto-applies.
• Conflict resolver — Scans for contradictions, emits recommendation (review_only or apply_via_safe_reconsolidation), never modifies memory directly.
• Reranker timeout & fallback — Configurable timeout (default 5s) with automatic fallback to vector-only ranking.
• schicht15 deduplication — KNOWLEDGE.md promotions tracked per workspace+agent. Double-promotion prevented via memoryId + optional contentHash.
• Obsidian bridge apply mode (safe) — mode: "apply" creates per-file backups, manifest (beforeHash/afterHash), and audit-log entry. Vault path confirmation required before first write.
• Rate-limited background jobs — Daily consolidation capped at 1×/day/agent; REM dreaming at 1×/week.

Recall boosters (additive)

These features run after normal recall and only append results; they never replace the primary recall result and never write memory data.

Semantic Lens

Reads a precomputed .plur1bus/semantic-lens-index.json from the workspace and adds a small number of community/bridge/faded memories that normal recall may have missed.

• Default: enabled: false in schema.
• Caps: maxLensMemories: 3, maxBridgeMemories: 2, maxFadedMemories: 1, maxCommunities: 2.
• Hard timeout: 50 ms; fallback returns base recall unchanged.
• No live graph recompute, no second recall path, no writes.

Conversation Reactivation Recall (CRR)

MVP reactivation hook that appends a <memory-reactivation> block when a conversation appears to resume after an idle gap, a compaction, or a continuation signal.

• Default: enabled: false in schema; visibleHints: false.
• Triggers: idle threshold (45 min), continuation signal, first substantive message, or post-compaction gap.
• Caps: maxReactivationMemories: 3, maxFadedReactivationMemories: 1, maxOpenThreads: 3, maxCommunities: 2.
• Hard timeout: 50 ms; silent fallback on error.
• State is module-level in-memory only; no writes to cards, tags, graph links, records, or quarantine.

Graph-link managed blocks / semanticDiscovery

Record notes can contain an idempotent managed block (id="graph-links") with wikilink edges. The block is regenerated, not appended, and conflicts with manual edits are reported.

• Tiers: explicit (memoryIds/sourceRefs), type (type-based rules), semantic (precomputed link index).
• Default semantic threshold: 0.78.
• semanticDiscovery builds .plur1bus/link-index.json from memory mirrors + vectors behind a confirmation gate; it is not auto-applied.

Technical frontmatter tags

Memory mirrors use technical filter tags, not semantic memory tags:

• plur1bus/memory
• plur1bus/agent/<id>
• plur1bus/workspace/<id>
• plur1bus/category/<cat>
• plur1bus/scope/<scope>

These tags are used for vault filtering and graph grouping; they do not carry semantic memory content.

User Commands

/plur1bus subcommands

Afterthoughts (delayed follow-ups)

When the last conversation ended 30–120 minutes ago with an open outcome (the user asked for details, or the topic was dropped mid-thread), the plugin can compose a short, casual follow-up message ("Mir ist zu … noch eingefallen…"). This is gated by the shared proactive governor budget, capped at one per day, and skipped for any topic already surfaced as an open thread today. Recommended cron: every 3 hours, run the exact command /plur1bus internal afterthought with announce delivery. The plugin command returns either a validated reply payload or OpenClaw's NO_REPLY suppression token. PLUR1BUS submits the exact allowlisted command through OpenClaw's public gateway-runtime dispatcher, so OpenClaw owns status finalization and at-most-once announce delivery without an outer carrier-model run. If the required native capability is unavailable, provisioning fails closed and leaves the feature job inactive. Custom prompts, surrounding whitespace, prefixes, and suffixes are never claimed.

Setting this cron up is automatic when its raw feature gates are explicitly enabled — see below.

Multi-agent feature-cron automation

node scripts/setup-feature-crons.mjs verifies the public native command-dispatch capability first. When healthy, it loads exactly one validated configuration snapshot with openclaw gateway call config.get --json, discovers bound agents, and idempotently plans up to ten jobs per agent plus one install-wide GC job. It fails closed without normal cron planning when the capability is absent, the gateway call fails, JSON is invalid, valid !== true, or sourceConfig/runtimeConfig is not a plain object. Custom prompts and unrelated jobs remain untouched. It never falls back to local config files or alternate raw/resolved fields.

The capability probe allows the host CLI 30 seconds per help call (7.8.0); a booting host with many plugins needs more than the former five. Ownership is by exact command: an operator-named job that runs /plur1bus internal <feature> for an agent is that agent's job, and only PLUR1BUS-named jobs are eligible for schedule migration. A non-delivery feature whose only owned jobs are disabled gets its best candidate re-enabled; the singleton collector is satisfied by a job on any agent; when openclaw agents list flags no default agent, main is treated as the default.

The two configuration views have separate roles: sourceConfig alone controls explicit raw feature gates and the raw skillMiner schedule; runtimeConfig alone controls effective bindings, accounts, and delivery. Runtime defaults cannot enable jobs. The eligible jobs are:

• persona-evolve: personaVoice.enabled && skillMiner.enabled; daily 04:15 local time (7.12.38; before: Sunday), staggered five minutes per agent, overridable via personaVoice.cron/personaVoice.timezone; no delivery. The shipped weekly slots are migrated to daily on the next setup run; operator schedules stay.
• afterthought: afterthought.enabled && (skillMiner.enabled || merging.enabled); every 3 hours; exact-command announce delivery with a direct text/NO_REPLY result.
• consolidate-daily: dailyConsolidation.enabled; daily 04:00 in Europe/Berlin; no delivery.
• classify-recent: criticalPush.enabled; every 3 hours; safe announce delivery of approved pushes or NO_REPLY.
• rem-dream: merging.enabled; daily 01:15 in Europe/Berlin; no delivery.
• skill-miner: skillMiner.enabled; raw Croner-compatible cron/timezone after conservative syntax validation, defaulting to Sunday 03:00 in Europe/Berlin (timezone: null means local time). Invalid literals, descending ranges (including named month/day ranges), names, modifiers, or literal-step forms are ineligible.
• discover-semantic-links: obsidianBridge.enabled && obsidianBridge.graphLinks.semanticDiscovery.enabled; daily 02:00 in Europe/Berlin; no delivery.
• gc-run: gc.enabled; daily 04:45 in Europe/Berlin, after consolidate-daily has produced the candidates; no delivery. This one is a singleton — runGcJob iterates over every agent database itself, so exactly one job is planned regardless of how many agents the install has.

Every job runs with --agent <agentId> --session isolated. Provisioning does not set model, fallback, token, auth, API, or other credential overrides, so OpenClaw's default LLM and per-agent credentials remain authoritative. The script remains idempotent and exit-0 for install safety, so it can run from any of these channels:

• npm install/npm postinstall — fires when the plugin is installed via npm install (e.g. npm install -g @cyb3rb1ade/plur1bus-memory).
• Gateway startup (deferred bootstrap) — runtime registration verifies the public Gateway method, CLI, and command dispatcher exactly once. A gateway_start handler schedules a bounded reconciliation after the gateway becomes reachable. Missing capabilities or invalid config leave exact owned feature jobs inactive; retries use bounded backoff. Disable provisioning with "featureCronSetup": { "auto": false }.
• Manual — /plur1bus setup crons (optionally --agent <id>/--account <acct> to force single-agent mode).

The /plur1bus doctor and /plur1bus status feature-cron hint is condition-derived, not "have we shown this before": it reads the marker file and only surfaces a hint when setup has never run, ran under an older plugin version, or ran but couldn't create everything it planned (some crons are still pending — e.g. no delivery target could be derived). It's silent once a current-version run reports nothing left to create.

• Bound-agent rule: only agents with bindings > 0 (i.e. an actual chat channel routes to them) get feature crons. Subagents (bindings === 0 — researchers, deep-divers, and other internal-use-only agents) are deliberately excluded; they have no chat to receive an automatic persona evolution or an afterthought delivery, and running these jobs against them would be pure compute waste.
• One agent per workspace: PLUR1BUS state for these jobs (persona voice, proactive-governor budget, afterthought dedup state, …) is keyed by workspace directory, not agent id. If two bound agents share a workspace, only one gets the crons (tiebreak: isDefault first, then most bindings, then alphabetically-first id) to avoid two crons double-firing against the same state files.
• Per-agent identity: all per-agent canonical names use plur1bus <feature> <agentId>. An existing job is owned only by an exact, case-sensitive agent id plus either its exact canonical name or exact first command line; missing or different agents are untouched. Every exact owned duplicate is inspected and reconciled, even when another duplicate is already safe.
• Safe delivery: outbound targets never come from allowFrom. Delivery-required jobs use only a conservatively validated Telegram binding match.peer.id (including t.me/<handle>) or effective account/root defaultTo. Every relevant non-ACP binding must agree on channel, and an account inherits only when match.accountId is truly absent. Omitted accounts resolve in order from an explicit valid defaultAccount, accounts.default, one sole named account, or a root account proven by configured botToken/tokenFile; routing fields alone never invent a root default account. Unsupported providers, wildcard, placeholder, redaction, zero-id, disabled-account, explicit empty/missing account, mixed-account, and conflicting target/channel/account states are rejected. Existing delivery seeds require exact mode: "announce"; case or whitespace variants are unsafe. A job without a validated target is created disabled with --no-deliver; every unsafe owned delivery job is disabled and stripped of delivery. Non-delivery jobs retain only missing delivery or exact mode: "none"; every other delivery object is removed.
• Agent discovery/input failure: if openclaw agents list --json fails, is unparseable, or yields no bound agents, no cron is mutated. Passing a validated --agent <id> forces one explicit agent; missing, option-like, or invalid --agent/--account values fail closed, and --account without --agent is rejected.

Installation

PLUR1BUS 7.7.1 requires Node.js 22.22 or newer and OpenClaw 2026.8.1 or newer. On an older host the installer refuses the package instead of deploying it: requires plugin API >=2026.8.1, but this OpenClaw runtime exposes <version>.

Install the published release through OpenClaw's package installer:

openclaw plugins install clawhub:@cyb3rb1ade/plur1bus-memory@7.5.3 \
  --acknowledge-clawhub-risk --pin

The same release is on the npm-compatible registry, if your @cyb3rb1ade scope already points there:

openclaw plugins install @cyb3rb1ade/plur1bus-memory@7.5.3 --pin

Or install the immutable GitHub Release tarball:

openclaw plugins install \
  https://github.com/Cyb3rb1ade/openclaw-plur1bus-memory/releases/download/v7.5.3/cyb3rb1ade-plur1bus-memory-7.5.3.tgz

To build from this source checkout instead, produce a tarball and install that artifact rather than linking the directory; a source link is not an equivalent package-compatibility test:

npm ci
npm test
npm pack
openclaw plugins install \
  npm-pack:/absolute/path/cyb3rb1ade-plur1bus-memory-7.5.3.tgz --force

Record the tarball's SHA-256 before transferring it. PLUR1BUS 7.5.3 never patches OpenClaw runtime files. Existing release artifacts remain unchanged and must not be relabelled as 7.5.0.

Restart the gateway after installing, so the new plugin version is loaded.

Then add a plugins.entries["memory-lancedb-namespaced"] block to your openclaw.json (see below).

Configuration

Minimal config block in openclaw.json. This is an explicit override example: the named gpt-4o-mini value and its credential are illustrative user choices, not PLUR1BUS defaults.

{
  "skills": {
    "workshop": {
      "autonomous": { "mode": "propose" }
    }
  },
  "plugins": {
    "slots": {
      "memory": "memory-lancedb-namespaced"
    },
    "entries": {
      "memory-lancedb-namespaced": {
        "enabled": true,
        "hooks": {
          "allowConversationAccess": true
        },
        "config": {
          "baseDbPath": "~/.openclaw/memory/lancedb-namespaced",
          "obsidianBridge": {
            "enabled": true,
            "mode": "augment",
            "vaultPath": "~/.openclaw/vault",
            "backupBeforeApply": true,
            "auditLog": true,
            "requireVaultPathConfirmation": true
          },
          "emotion": {
            "tier": "auto",
            "t2": { "enabled": true },
            "t3": {
              "enabled": true,
              "model": "gpt-4o-mini",
              "apiKey": "${OPENAI_API_KEY}",
              "escalationConfidence": 0.85,
              "timeoutMs": 4000
            },
            "moodInfluence": 0.3,
            "intensityHalfLifeFactor": 1.0,
            "temperaments": {
              "bernhardine": { "preset": "warm", "baseline": { "joy": 0.35, "trust": 0.5 }, "sensitivity": 1.5, "decayMultiplier": 1.3 }
            }
          },
          "dailyConsolidation": {
            "enabled": true
          },
          "dreaming": {
            "enabled": false,
            "narrative": { "enabled": true, "storeAsMemory": true }
          },
          "merging": {
            "enabled": true,
            "mode": "safe-versioned",
            "autoApply": false
          },
          "reranker": {
            "enabled": true,
            "timeoutMs": 5000,
            "fallbackOnError": true
          },
          "security": {
            "allowChatConfigCommands": true,
            "allowModelDestructiveMemoryOps": true,
            "allowedUserIds": [],
            "allowedChatIds": []
          },
          "runtime": {
            "embeddingCacheEnabled": true,
            "embeddingCacheMaxEntries": 128,
            "embeddingCacheTtlMs": 300000,
            "embeddingCacheScope": "agent",
            "llmResultCacheEnabled": true,
            "llmResultCacheTtlMs": 86400000,
            "llmResultCacheMaxEntries": 256,
            "llmResultCachePersist": false,
            "llmResultCacheMaxBytes": 67108864,
            "llmResultCacheMetrics": true
          }
        }
      }
    }
  }
}

dreaming.enabled: false is the OpenClaw memory-core sidecar gate, not the PLUR1BUS narrative toggle. Keep it false when PLUR1BUS owns consolidation or REM; otherwise OpenClaw also loads memory-core dreaming. When skillMiner is enabled, use Skill Workshop autonomous propose or off so both learning systems share the governed proposal queue without independently applying overlapping skills.

hooks.allowConversationAccess: true is mandatory for this trusted memory plugin. It authorizes the official typed before_agent_reply hook used for automatic recall and for the fail-closed admission boundary of exact feature commands; OpenClaw otherwise withholds the conversation body. The installer enforces this single permission even in preserve mode while keeping all unrelated hook and feature choices unchanged.

All paths default to $HOME/.openclaw/... if omitted. OPENCLAW_CONFIG_PATH and OPENCLAW_HOME env vars override the lookup of the gateway config file used by the toggle commands.

Choosing embedding and reranking models

PLUR1BUS ships a keyless fallback for both stages and recommends something stronger for each. The two choices are not equal in weight: the embedding model decides how well recall separates hits from noise, the reranker only refines the candidates the embedding already found.

Embedding. The fallback intfloat/multilingual-e5-small (fixed 384 dimensions) packs its vectors into a narrow cone. Texts that have nothing to do with each other still land at a cosine similarity around 0.84, which is a recall score of roughly 0.86 on the 1 / (1 + distance) scale PLUR1BUS uses. Hits and background noise sit so close together that neither the ranking nor a threshold can separate them cleanly, and the band 0.96 to 1.0 reserved for agent-marked cards as well as the duplicate threshold of 0.95 lose their meaning. This is exactly what the compatibility lab showed when it measured with E5 and no reranker. The recommended model spreads similarities much wider: jinaai/jina-embeddings-v3 (multilingual, Matryoshka dimensions from 32 to 1024, CC BY-NC 4.0 license consent required). With it, ranking and thresholds do their job. Since 7.12.0 the installer's first option is Jina v5 Text Nano (see the lab test below); OpenAI is the hosted alternative, Jina v3 stays selectable for existing installs, and E5 is only the small keyless fallback. The embedding model is switched from the PLUR1BUS tab as well (controlUi.writeActions: "all"): the button picks the target model, model preparation downloads and verifies it, and the re-embedding migration then runs from the same page with a dry run, a copy, and a separate switch. Unlike the reranker switch, every card is embedded again along the way, so it takes time and disk space, and the old generation stays available for rollback.

Which dimension. Jina v3 is trained with Matryoshka representation learning, so the vector can be cut to 32, 64, 128, 256, 512, 768 or 1024 dimensions. The Jina paper (table 7) reports retrieval nDCG@10 of 63.35 at 1024, 63.30 at 768, 63.16 at 512, 62.72 at 256, 61.64 at 128, 58.54 at 64 and 52.54 at 32. Keep the PLUR1BUS default of 1024: at memory scale the vectors are small anyway (50,000 cards are about 200 MB of float32 vectors), and the duplicate threshold of 0.95 and the reserved band above 0.96 are calibrated on full-width similarities. 512 is the economical choice for very large stores or weak hardware; it halves storage and ANN cost for a loss of 0.2 points. Do not go below 256: from 128 down the recall loss becomes visible and the similarity spread compresses, which moves every threshold. The dimension is baked into the table, so changing it later means another re-embedding run.

Jina v5 Text Nano (7.11.0), an option with the lab test pending. The EuroBERT-based nano model (239M parameters, 12 layers, 768 dimensions with Matryoshka down to 32, 15 European languages including German, CC BY-NC 4.0) runs as the upstream Q8 export at roughly a quarter of v3's compute and about 265 MB on disk; on the reference machine one card embeds in about 0.45 s and the process grows by about 120 MB. Jina reports MMTEB 65.5 against 64.44 for v3. Queries and documents are told apart by the Query: and Document: prefixes. Whether it becomes the proposal for new installs is decided by the PLUR1BUS lab test on real memory cards, not by the version number. That test ran on 5 September 2026 (OpenClaw 2026.9.1, Transformers.js 4.2.0, 12 cores shared with other work): 1,031 active cards of one agent, 72 real recall queries from the gateway log, the production ranking (OpenAI text-embedding-3-large, 3,072 dimensions) as reference, both models through the real provider with the 512-token cap, in separate processes, back to back.

Reading it: on ranking quality the two are level within the noise of 72 queries. Nano separates hits from the noise band much more clearly (v3 compresses similarities into a narrow band around 0.54, which is what makes thresholds hard to set), embeds cards three times faster in batches, loads in a third of the time and needs half the memory; v3 answers a single query twice as fast, which at a quarter of a second is not a user-visible difference. Both models push more distinct pairs above the 0.95 duplicate threshold than the production model, mostly genuine near-duplicates. On these numbers the maintainer made Nano the proposal for new installs in 7.12.0 and recommends existing installs the migration from the dashboard: prepare the Nano target in Model Preparation, run the dry run, copy, then switch; the old generation stays for rollback. Nothing migrates on its own.

Reranking. The reranker reviews the 40 candidates the vector search returns and removes the ANN noise. Two local models are pinned, both quantized ONNX through Transformers.js: bge-reranker-v2-m3, which the installer proposes and the runtime uses when nothing else is configured, and jinaai/jina-reranker-v2-base-multilingual, available under the wizard's advanced options and as a dashboard switch. Switching between them is a runtime choice without any data migration (see controlUi.writeActions). The two models differ in size and in what they are good at:

Where two numbers are given, the first comes from the Jina model card and the second from the jina-reranker-v3 paper; the two evaluations use different candidate sets, so compare within a column pair, not across. On English and on the 26-language MKQA set the two are level. BGE is clearly ahead on MIRACL, the broad multilingual retrieval benchmark, and slightly ahead on tool retrieval. Jina is clearly ahead on long documents and on code, and at half the depth it reranks noticeably faster on CPU; Jina quotes up to 15x the document throughput of BGE with flash attention on a GPU. For PLUR1BUS memory cards, which are short German or English summaries, multilingual ranking quality is what counts, and long documents and code do not occur. The recommendation is therefore BGE: ahead or level on the benchmarks that matter here, Apache 2.0 without a license consent, at the price of roughly twice the CPU time per query, which is a fraction of a second for 40 short candidates on an ordinary server. Pick Jina only when a weak CPU runs into the reranker timeout of five seconds; the quality loss is small.

Sources: Jina model card, BGE model card, jina-reranker-v3 paper, table 2, Jina Reranker v2 announcement.

Settings for both stages are described in docs/configuration.md; the installer offers the preparation profiles for the local models.

Named storage namespaces

Omitting namespaces preserves the legacy-flat layout exactly: {baseDbPath}/{agentId}. Named routing is enabled only by supplying the strict object explicitly:

{
  "baseDbPath": "~/.openclaw/memory",
  "namespaces": {
    "activeWriteNamespace": "lancedb-local",
    "activeRecallNamespaces": ["lancedb-local"],
    "legacyReadOnlyNamespaces": ["lancedb-namespaced"],
    "crossNamespaceRecall": true
  }
}

An explicit baseDbPath may be the named root, as above, or the active writer leaf (~/.openclaw/memory/lancedb-local); both forms resolve to the same layout. Namespace identifiers must match ^[A-Za-z0-9][A-Za-z0-9_-]{0,63}$. The writer must occur in active recall, and active namespaces must be disjoint from legacy read-only namespaces. Legacy namespaces participate only when crossNamespaceRecall is exactly true; they are opened without table creation, schema migration, or mutation.

Every recalled table must use the configured embedding dimensions. Each table runs the existing recall pipeline. When multiple live tables participate, PLUR1BUS waits for all of them and performs one stable global score merge; the one-table path remains direct. Duplicate IDs and canonical heading/text are collapsed, canonical plus memory results share the configured output cap, and child decision traces are replayed through the existing trace caps. A namespace changes storage routing for the current agent only. Sharing between agents, workspaces, or users is separate ACL work owned by B13.

OpenClaw chat-LLM routing

Chat models are selected per owning feature. If an optional feature model is absent, PLUR1BUS uses the effective OpenClaw agent model and sends no model property. Features never inherit merging.model, its endpoint, credential, or headers. Existing feature/profile activation, budgets, confirmation gates, rate limits, and fail-soft behavior remain unchanged; Safe produces zero PLUR1BUS native/direct chat calls.

The four selection modes are openclaw-default (native with no model), openclaw-override (feature-local model through OpenClaw), direct-override (feature-local model plus direct transport), and unavailable. failed is the stable diagnostic outcome when a selected transport rejects. Provider/model metadata returned by OpenClaw may be recorded without credentials, prompts, or headers. Native routes bypass the PLUR1BUS result cache; complete direct routes retain exact caching.

Direct transport without a feature-local model fails closed and sends no request. A configured credential that is unresolved is unavailable; it never falls through to native OpenClaw host credentials and does not abort plugin registration. runtime.llm.complete missing or unavailable is fail-soft and does not select a hard-coded model.

A session-bound command capability omits agentId. Global hook, tool, and background calls retain the target agent and require entry-level llm.allowAgentIdOverride:true. A model-only native override requires llm.allowModelOverride:true and obeys allowedModels. Installer preserve never grants LLM trust, and neither Safe nor Recommended adds those entry-level bits.

runtime.llm.complete resolves the effective primary selection and does not execute the configured model fallback array in the installed runtime. PLUR1BUS neither claims nor emulates a host fallback chain.

LLM result cache

PLUR1BUS caches only exact, agent-scoped results from an explicit allowlist of deterministic internal LLM transformations. The default in-memory cache uses a 24-hour absolute TTL (llmResultCacheTtlMs: 86400000, clamped to 60 s–7 d) and holds 256 entries per plugin registration (llmResultCacheMaxEntries, clamped to at most 10,000). Optional prompt-free SQLite persistence is off by default; when enabled with llmResultCachePersist, it stores hashed keys, results, usage metadata, and timestamps under the memory database path without storing plaintext prompts, credentials, or headers. llmResultCacheMaxBytes defaults to 67,108,864 bytes and is clamped to at most 1 GiB; clamped values log a warning.

The six runtime settings are llmResultCacheEnabled (default true), llmResultCacheTtlMs (default 86400000), llmResultCacheMaxEntries (default 256), llmResultCachePersist (default false), llmResultCacheMaxBytes (default 67108864), and llmResultCacheMetrics (default true). Missing values come from the manifest; an explicit false remains authoritative.

Operational notes:

• Persistence uses the built-in node:sqlite module available throughout the supported Node.js runtime range; if SQLite initialization is unavailable, the cache falls back to memory-only.
• Persistence stores LLM response text as plaintext (directory 0o700, file 0o600 under the memory database path). Responses may contain condensed memory content — enable persistence only where that is acceptable.
• Integrated call sites send temperature: 0 for determinism, and llm-call.js now actually forwards temperature to the provider (previously the setting was silently ignored). Existing configs that set temperature therefore change their effective provider behavior.

Non-goals and bypasses:

• Ordinary main-chat/model responses and calls with a missing or unknown cache purpose always remain live. That includes weather requests such as wie wird das Wetter morgen?.
• Live or creative wiki, critical classifier/push, dream narrative, dream echo, afterthought, persona voice, and overlay paths are not cached.
• Direct Tier-3 API-client calls and emotion calls without a real agent scope are not cached.
• The cache does not perform semantic matching, share results across agents, cache rejected upstream calls, or replace provider-side prompt caching.

The /state status section reports cache hit rate, memory/persistent hits, persistence state, and avoided input/output tokens. It intentionally reports token counts, not money.

emotion.t3 — the tier-3 emotion classifier uses the effective OpenClaw agent model when its model is absent. A complete feature-local direct override may instead provide its own model and transport. If neither route is available, the classifier falls back to Tier-2 heuristics: it does not label cards, so it never poisons results by marking everything fakt.

emotion.temperaments — per-agent emotional temperament. Ships with generic defaults only (main slightly more sensitive, everyone else balanced). Pick a preset via /plur1bus temperament <preset> (ausgewogen, warm, kühl, feurig, stoisch) — requires a gateway restart. Mood always derives from conversation content; the temperament only shapes how strongly and how long it swings. The current mood is written to .emotional-state.json (machine-readable, survives restarts) and .current-mood.txt (human-readable) in the agent workspace, injected as a mood line into the recall context, stamped on every memory card (moodContextAtCapture), and emotionally intense memories decay slower (intensityHalfLifeFactor).

security.allowedUserIds / security.allowedChatIds — identity-aware authorization for commands and destructive flows.

• If both lists are empty, non-destructive commands can run in private 1:1 contexts; destructive commands are denied in groups/unknown channels.
• If either list is configured, destructive commands require userId membership in allowedUserIds (chatId alone is never sufficient), plus allowedChatIds when that list exists.
• Whitelists remain stable with /enable, /disable, /plur1bus setup, /forget, /correct and confirmation flows.

security.allowChatConfigCommands (default true) — disables operator-level config mutating commands (/enable, /disable, /plur1bus setup) when set to false. Use this in shared channels if you want a hard stop on chat-driven writes. Writes are still guarded by a file lock.

security.allowModelDestructiveMemoryOps (default true) — keeps model-facing tools memory_forget and knowledge_update available unless you explicitly disable them.

Scope-sichere Speicherung

scope values now support agent-private (default), workspace and user.

• agent-private remains per-agent.
• workspace shares by workspace.
• user is owner-bound: der aufrufende userId wird gespeichert und bei Sichtbarkeit/Mutation geprüft.

Freigegebene Memory-Pools (B13)

/share <id> kopiert eine sichtbare Karte nach bestätigter, an Benutzer und Chat gebundener Bestätigung in den Workspace-Pool. /share <id> --user nutzt dieselbe Bestätigung, erzeugt aber einen nur für denselben Kanal, Account und Benutzer sichtbaren User-Pool. Die Grammatik ist strikt: nur ein vollständiges UUID-id, optional genau --user, oder /share confirm <nonce> sind gültig; unbekannte oder doppelte Optionen werden vor jedem Store-, DB-, Embedding- oder Provider-Zugriff abgelehnt. Eine Freigabe ist copy, never move: die private Ursprungskarte bleibt unverändert, und die autorisierte Shared-Kopie enthält einen kanonischen Origin-Verweis. Recall darf die optionalen Shared-Quellen ergänzend lesen und dedupliziert den kanonischen Ursprung; sie ersetzen weder primären Recall noch dessen ACL.

Physische Routen sind kein benutzergesteuerter Pfad: ihre Segmente sind höchstens 64 Zeichen lang und werden als .plur1bus-shared/workspaces/w-<62hex> oder .plur1bus-shared/users/u-<62hex> abgelegt. Die Berechtigung bindet den kanonischen Workspace konfliktablehnend (keine versteckte Alias-Priorität) und den vollständigen Kanal+Account+Benutzer-Prinzipal. Fehlende oder abweichende Bindungen sind nicht sichtbar und nicht mutierbar; fehlend und verweigert werden gleich behandelt. /memory und /share --user verwenden den direkt vom Host gelieferten Account.

Der aktuelle OpenClaw-Hook kann die optionale automatische User-Shared-Recall Quelle ausschließlich bei aktiviertem autoRecall verwenden. Er benötigt einen account-tragenden Session-Key, ein exaktes Host-Run-Ticket oder eine konservative default-only Account-Topologie. Native und Slash-Kommandos prägen absichtlich kein Route-Ticket, weil behandelte Kommandos den Prompt-Hook nicht erreichen. Bei mehrdeutigen benannten/multi-account Main-, Group- oder Channel-Turns wird nur diese optionale Quelle ausgelassen; andere Recall-Quellen bleiben unberührt. Ein zuletzt gespeicherter Session-Route-Wert ist kein turn-gebundener Account-Beweis.

Legacy rows that used the old workspace_shared scope remain in their authoritative private table until an operator explicitly migrates them. Start with the non-mutating audit:

/plur1bus migrate-legacy-shared

Use --report <name.json> for a fixed private report name, and resume a bounded dry run with the opaque --cursor <token> returned by the previous run. After reviewing the report, run --apply without a dry-run cursor; apply re-reads each source row, writes and verifies an idempotent workspace copy, and only then marks the legacy source. The command never deletes or re-scopes the source row: workspace_shared legacy rows are not reinterpreted. The operation is bounded per run to 250 rows, 4 MiB source bytes, 100 provider calls, and 60 seconds. The opaque cursor pins source versions and dry-run mode; an unavailable or changed pinned version, mode mismatch, checksum/binding failure, timeout, or uncertain commit aborts the run and requires the documented continuation or a restart without the cursor. Apply never accepts a dry-run cursor. It is operator-destructive, so it requires the same user authorization as /forget; cron identity does not bypass that gate. Reports are no-clobber 0600 JSON files below .plur1bus/migrations/ and exclude memory content, vectors, evidence, and provenance.

The migration runs only through the destructively authorized initialized runtime command; there is no standalone DB/config/credential bootstrap. Multi-Namespace, Neo/Obsidian aliases, Semantic Lens, CRR, the OpenClaw default LLM, and per-agent credentials do not change under sharing or migration.

security.allowModelDestructiveMemoryOps (default true) — the model-facing tools memory_forget and knowledge_update mutate persistent memory/knowledge state. Set this flag to false if you want a hard opt-out for model-driven destructive memory writes.

Feature profiles

Core memory (capture, recall, search) works from manifest-safe defaults without profile confirmation. Argument-less setup only lists the choices, and start is read-only status/onboarding guidance:

# In Telegram
/plur1bus setup
/plur1bus start

Apply a profile only by naming it explicitly:

/plur1bus setup safe
/plur1bus setup recommended

An explicit selection records setupProfile and featuresConfirmedAt. Recommended enables additional features while retaining merge and Obsidian safety gates; vault discovery alone never counts as confirmation.

Architecture

LanceDB is the authoritative store: every memory card lives there first, indexed per agent for isolation. The Obsidian bridge mirrors cards into a Markdown vault so the user can read, link, and edit them with normal tools; LanceDB stays the source of truth and the bridge re-syncs on changes.

A daily consolidation job detects duplicates and generates merge proposals (never auto-applies). A critical-push classifier (run via the OpenClaw-managed cron as the exact command /plur1bus internal classify-recent) labels recently captured cards by sensitive entity type (person, relationship, birthday, money/account, health, access/password) using the configured chat model, and — when a per-agent daily threshold (maxPerDay) is not yet exceeded — emits a short confirmation message per critical card. The command handler converts returned pushMessages into a validated native command reply; OpenClaw's dispatcher owns finalization and delivery. Multiple push texts are combined in their original order; partial classifier failures are reported alongside successfully produced pushes. The per-day counter is enforced across runs, and each card is classified exactly once, so no card is pushed twice.

The recall pipeline runs embedding → LanceDB vector search → optional query refinement → temporal filter → canonical KNOWLEDGE.md search → score/status processing → graph spread and hydration → budget allocation → optional rerank → deduplication → ACL filtering → finalization. The caller may then append bounded Semantic Lens and Conversation Reactivation Recall results; neither replaces the primary recall.

Development

npm install
npm test              # full serialized Node test runner

No build step. ESM-only. Tests are unit-level and DB-free; the LanceDB adapter is mocked behind a thin interface.

Recall safety in v6

Recalled memories are rendered as historical evidence, not as current user requests. A memory that contains an old imperative such as a download, send, write, delete, install, purchase, network action, or command must not trigger that action unless the current visible user turn asks for the same action.

The recall block uses escaped metadata attributes and wraps recalled text in quoted-evidence elements, so prompt boundaries stay explicit even when old memory text contains tool-like markup.

Migration from 5.x

Version 6.x is a major upgrade. If you ran 5.x:

• Schema migration — LanceDB table schema is auto-migrated on first init(). New columns: status, versionNumber, previousVersion, supersededBy, updateSource, updateEvidence, reconsolidationConfidence, versionCreatedAt, updatedAt. Migration is idempotent and non-destructive.
• Explicit profile selection — Missing values use manifest-safe defaults. Use /plur1bus setup safe or /plur1bus setup recommended only when you intentionally want to persist a profile; /plur1bus setup and /plur1bus start are non-mutating.
• Merging is proposal-only — merging.autoApply defaults to false. Merge candidates are written to merge-proposals.jsonl instead of being applied automatically. Set autoApply: true to restore 5.x behavior.
• Obsidian bridge apply mode — New mode: "apply" with safety gates (backups, audit log, vault path confirmation). Default is mode: "augment" (read-only). Confirm vault path explicitly before first write.
• Command input handling — Hard length limits removed. Very long inputs are semantically compressed; beyond 100k chars use a file or vault source.
• Config keys added — reranker.timeoutMs, reranker.fallbackOnError, merging.autoApply, merging.mode, obsidianBridge.backupBeforeApply, obsidianBridge.auditLog, obsidianBridge.requireVaultPathConfirmation, obsidianBridge.morningReview.status, obsidianBridge.eveningReview.status, emotion.tier, emotion.t2.enabled, emotion.t3.enabled, emotion.t3.model, emotion.t3.apiKey, emotion.t3.escalationConfidence, emotion.t3.timeoutMs, emotion.moodInfluence, emotion.intensityHalfLifeFactor, emotion.temperaments.<agentId>.

See v5_TO_v6_MIGRATION.md for the full migration guide.

License

MIT — see LICENSE.