Synapse

Don't sync memory across machines — link it.

Most tools that keep Claude Code's memory consistent across machines sync it: copy files back and forth, merge on conflict, hope nothing diverges. Synapse takes a different position — there is exactly one physical copy of your memory (a git repo you own), and every machine, every project, references it through a filesystem link (an NTFS junction on Windows, a symlink elsewhere). Nothing to merge, because nothing is ever duplicated in the first place.

                        ┌───────────────────────┐
                        │   hub — a git repo     │
                        │   you own, ONE copy    │
                        └───────────┬───────────┘
                        junction / symlink, never a copy
              ┌─────────────────────┼─────────────────────┐
      ┌───────▼───────┐     ┌───────▼───────┐     ┌───────▼───────┐
      │ machine A      │     │ machine B      │     │ machine A      │
      │ project 1      │     │ project 1      │     │ project 2      │
      │ .claude/memory │     │ .claude/memory │     │ .claude/memory │
      └───────────────┘     └───────────────┘     └───────────────┘

How it works

1. You point Synapse at a private git repo — the hub. It gets cloned once per machine.
2. Every project that wants memory gets a link (<project>/.claude/memory by default) pointing straight at that one clone. Read or write through the link, and you're reading or writing the hub directly — no sync step in between.
3. A SessionStart hook keeps a local RAG index fresh (semantic + exact-match search) when the hub changed since last time. A SessionEnd hook commits and pushes automatically, gated by a secret scan — nothing with a shape resembling a credential ever gets committed silently.

That's the whole mechanism. Everything else (linting the memory, watching for stale links, scanning for merge candidates, tracking the competitive landscape) is built on top of it, never a second source of truth.

Why link instead of sync

See docs/DESIGN.md for the full reasoning and trade-offs, not just the pitch.

Commands

Install

/plugin marketplace add smaurier/claude-synapse
/plugin install synapse@synapse

(This repo doubles as its own marketplace — no separate listing to submit to.) Dev mode also works, without registering a marketplace:

claude --plugin-dir /path/to/claude-synapse

Then run /synapse-init and give it the URL of a private git repo to use as your hub (Synapse refuses to link a public one on GitHub — memory doesn't belong in a public repo). Requires Node >= 22.5 (the local search index uses node:sqlite, a builtin only available from that version on).

Status

Core mechanism built and tested: 221 tests, real-hub validation (not just fixtures) surfaced and fixed several real bugs along the way — a lexical fallback for bare-acronym search, word- boundary matching to avoid false positives, incremental indexing, chunk-aware merge detection. Scale-tested too, not just assumed: search stays fast into the thousands of files; the one piece that didn't (merge-candidate detection, O(n²)) is guarded against a configurable file-count ceiling rather than left to hang. Self-hosted marketplace listing added — see Install above. /synapse-init still configures one project at a time by hand, but a root scanned once via /synapse-refresh-projects now stays covered automatically on every later periodic audit.

Learn more

docs/DESIGN.md covers the architecture decisions behind each piece, the trade-offs made on purpose, and the limitations documented as such rather than left implicit.

License

MIT