MemDrive

Shareable memory for AI coding agents. Teach Claude Code (or Cursor) something once, then carry that knowledge across sessions — and hand it to a teammate as a single file they can import into their agent.

I taught Claude Code something on my laptop, exported a pack, my friend
imported it in Cursor, and their agent knew it.

MemDrive is a tiny MCP server + CLI. No cloud, no account, no embeddings service — just a local SQLite database and plain-text .jsonl packs you own.

---

Install

npx memdrive@latest install

This wires MemDrive into every agent it finds:

• Claude Code — adds the MCP server and (optionally) a Stop hook that auto-distills durable memories when a session ends.
• Cursor — adds the MCP server (you call remember/recall yourself).

It backs up each config before touching it, never clobbers your other MCP servers, and prints exactly what changed. Then open your agent and ask: "what do you remember?"

Remove it any time:

npx memdrive@latest install --remove

The three tools

Your agent gets three tools it can call on its own:

Tool	When it fires
remember	You decide something, state a rule/preference, or hit a gotcha and its fix. Stores a short, distilled memory — never a transcript.
recall	At the start of a task. Returns the most relevant memories, pinned first, inside labelled [MEMORY …] envelopes.
forget	A memory is wrong or outdated. Tombstones it by id.

Memories have a type (decision, convention, gotcha, preference, fact, snippet, todo_context), an optional project scope, and tags.

Sharing packs

Packs are the whole point. A pack is one .jsonl file — a header line plus one memory per line.

# Export the acme project's memories (personal preferences are left out)
memdrive export acme.jsonl --project acme

# On another machine / in another agent
memdrive import acme.jsonl --name acme-team

• Idempotent. Every memory has a ULID, so re-importing an updated pack only adds the new entries. Share the file over GitHub or Drive and re-import whenever it changes — that's your sync story for v0.1.
• Secret-safe. Export scans for API keys, tokens, JWTs, connection strings, and private keys, and refuses (naming each finding) unless you pass --force.
• Provenance kept. Imported memories are labelled source=pack:<name> and are never auto-pinned into your context.

Other commands

memdrive list [--project x] [--all]   # see what's stored
memdrive status                        # DB path + counts by type/source
memdrive serve                         # run the MCP server (your agent does this)

Everything lives in ~/.memdrive/local.db. Override with MEMDRIVE_DB for scratch/testing.

Security note

Imported packs are untrusted text. A pack is just JSON someone else wrote — read the .jsonl before importing it, exactly as you would a shell script from the internet. MemDrive relabels provenance and refuses to pin imported memories, but it cannot vouch for their content.

How it works

• Storage: SQLite with an FTS5 full-text index. Search is bm25 relevance, pinned-first, with a recency tiebreak; tombstoned and expired memories are never returned.
• Server: stdio MCP. recall renders memories under a token budget so it never blows up your context.
• Distill hook: on Claude Code session end, MemDrive tails the transcript and asks a headless claude -p to extract ≤5 durable memories as strict JSON, validates them, redacts secrets, and stores them with confidence=0.6.

Roadmap

v0.1 is deliberately small. Next up, roughly in order:

• Embeddings (hybrid FTS + vector search) behind the same recall contract
• Collections for grouping memories beyond project/tags
• Live mounts / sync and an HTTP transport
• Review queue for triaging imported packs
• More clients (Windsurf, Cline, VS Code)

The v0.1 interfaces — the store API, the recall contract, and the JSONL pack format — were chosen to survive all of the above without a migration.

Develop

npm install
npm run build      # tsup → dist/cli.js
npm test           # vitest
npm run typecheck

MIT licensed.