Synapse Memory

A memory for your coding assistant. One file on your disk. Nothing to install, nothing to keep running, nothing sent anywhere.

Your assistant forgets everything between sessions. Synapse Memory gives it somewhere to look things up: past decisions, what worked, what broke. It hands back a short relevant answer with a source instead of a wall of text, and it knows when a fact was true, so an old note does not outrank a newer one.

It is one binary. Searches come back in well under a millisecond whether the file holds five notes or 294,000.

Core promise: best context, not biggest context.

---

Why Synapse Memory

---

Install

Option 1 — Prebuilt binary (recommended)

Download the latest release for your platform from GitHub Releases:

# Linux/macOS
tar -xzf synapse-<target>.tar.gz
cd synapse-<target>
sudo cp bin/synx bin/synapse-ultra /usr/local/bin/

# Windows
Expand-Archive synapse-x86_64-pc-windows-msvc.zip
# Copy bin\synx.exe and bin\synapse-ultra.exe to a PATH directory

Option 2 — Cargo install

cargo install --locked --path crates/synapse-cli
cargo install --locked --path crates/synapse-ultra

Option 3 — Build from source

git clone https://github.com/Supersynergy/synapse-agent-memory.git
cd synapse-agent-memory
cargo build --release -p synapse-cli -p synapse-ultra
# Binaries at target/release/synx and target/release/synapse-ultra

Verify

synx --version
synapse-ultra --version
synx doctor -f ~/.synapse/brain.db

---

5-line demo

synx -f ~/.synapse/brain.db prime .
synx -f ~/.synapse/brain.db remember --kind decision "Use Synapse context packs before major code edits."
synx -f ~/.synapse/brain.db context "current repo task" --mode coding
synx -f ~/.synapse/brain.db fresh-context --cwd . --prompt "latest package API changes"
synx -f ~/.synapse/brain.db doctor --fix

---

Production tools (new in v2.1.0)

Health check — 11-point audit

synapse-ultra health --db ~/.synapse/brain.db
synapse-ultra health --db ~/.synapse/brain.db --json   # for monitoring

Checks: integrity, WAL mode, synchronous, foreign keys, FTS5 index, indexes, triggers, schema version, DB size, page cache, ultra schema.

Backup — zstd-compressed with manifest

synapse-ultra backup --db ~/.synapse/brain.db
# → ~/.synapse/backups/brain-<ts>.db.zst (64% compression, sha256 manifest)

Metrics — Prometheus + JSON

synapse-ultra metrics --db ~/.synapse/brain.db --format prometheus
synapse-ultra metrics --db ~/.synapse/brain.db --format json

Exposes: synapse_events_total, synapse_decisions_total, synapse_docs_total, synapse_db_size_bytes, synapse_tags_total, synapse_tag_associations_total, and more.

Tag system — auto-rules, bulk-tagging, export/import

synapse-ultra tags add rust --color "#dea584" --description "Rust language"
synapse-ultra tags tag 42 rust --source manual
synapse-ultra tags bulk --ids 1,2,3,4,5 rust --source auto
synapse-ultra tags rule refactor refactoring      # keyword → tag auto-applied on ingest
synapse-ultra tags merge rust-lang into rust      # repoint + delete
synapse-ultra tags cleanup                         # remove orphan tags
synapse-ultra tags stats
synapse-ultra tags export > tags.json
synapse-ultra tags import tags.json

---

Architecture

graph TD
    CLI["synx CLI"] --> Daemon["synapsed (Unix-socket)"]
    MCP["synapse-mcp (agent memory tools)"] --> Daemon
    Daemon --> Core["synapse-core (SQLite + FTS5 + sqlite-vec)"]
    CLI --> Context["context / prime / fresh-context / feedback"]
    Context --> Core
    Context --> Learn["synapse-learn feedback loop"]
    Daemon --> Ann["synapse-ann (usearch HNSW)"]
    Daemon --> Fts["synapse-fts (Tantivy)"]
    Daemon --> Rerank["synapse-rerank (ColBERT-i8)"]
    Core --> DB[(brain.db)]
    Ann --> DB
    Fts --> DB
    Ultra["synapse-ultra (events + graph + tags + ops)"] --> DB

---

Crate map

Production (crates/)

Experimental (experimental/)

Stubs — excluded from default workspace build.

---

Benchmarks (verified, M4 Max)

Full bench files in bench-dashboard/.

ANN — SIFT-1M 128d (1M vectors, 1000 queries)

Source: SIFT1M_BENCH_2026-05-12.md

Hybrid search — production daemon (294k docs)

Source: REAL_BENCH_2026-05-11.md

Eval — LoCoMo + LongMemEval

Automated harness in eval/. Run:

python3 eval/harness.py download
python3 eval/harness.py ingest --db /tmp/eval-brain.db
python3 eval/harness.py run --db /tmp/eval-brain.db --k 5
python3 eval/harness.py report

Metrics: Recall@k, MRR, latency p50/p95, per-category breakdown.

Durability — SQLite-WAL

Source: FAIR_DURABILITY_BENCH_2026-05-13.md

---

Feature matrix

---

Roadmap

• Datalog semi-naive: delta-join + HashMap index (currently broken above 100 facts)
• HNSW parallel batch insert (target <10s for 1M vs current 197–672s)
• glass-backend CPU-SIMD beam search (expected ≥2× QPS vs usearch)
• io_uring durability bench on bare-metal Linux
• Metal dispatch verification for synapse-mlx-olap
• MTEB full 56-task suite (2/56 measured today)
• Python wheel publish to PyPI (synapse-py via maturin)
• synapse-raft production hardening

---

License

MIT — library crates. synapse-engine — source-available Engine License (non-commercial free; commercial license available).

See LICENSE-CORE.md and LICENSE-ENGINE.md.

Contributing

See CONTRIBUTING.md. Known issues: KNOWN-ISSUES.md.