An MCP memory server for coding agents. Stored facts get checked against
the filesystem and git before the model relies on them, so a memory
that has rotted is flagged instead of quoted back at you.
Install
Claude Code — two slash commands, zero config:
/plugin marketplace add 0Mattias/bettermemory
/plugin install bettermemory@bettermemory
Anything else that speaks MCP:
uv tool install bettermemory
bettermemory init --client claude-desktop
Restart the client. bettermemory doctor exits 0 when it's wired
correctly and prints a one-line fix for anything that isn't.
Try it without installing: uvx bettermemory try writes a memory
citing a file, deletes the file, and shows the next search flagging it.
Offline, throwaway store.
What it does
- Checks memory before believing it. Every hit carries a staleness
verdict: calendar age, whether the paths it cites still exist, the
commits landed since it was last confirmed, and how the record
entered the store (written here, pulled, or placed by hand).
Declared claims
(
path, path::symbol, path::NAME=literal) are re-checked
against the working tree; a claim that stops being true blocks the
stamp instead of riding along.
- Retrieval is deliberate. Memory is a tool call, not an
injection. Write gates bounce transient state, secret-shaped tokens
and near-duplicates; claims about you are labelled as inferences,
so they stay distinguishable and correctable.
- The code is the model. Search is deterministic lexical ranking —
keyword + BM25, fused — over your own vocabulary. No embedding
models, no downloads, nothing to warm up, same answer every time.
- Plain files. One markdown file per memory. Greppable,
git-syncable, no cloud, no account. The SQLite index beside the
files is a derived cache you can delete;
bettermemory reindex
rebuilds it.
- Rot gets acted on, not accumulated. Episodes journal per-session
run-state without polluting durable search; health telemetry and
curation tools surface what drifted, what went cold, and what
contradicts what.
- Receipts, not adjectives. The claims above are measured by
preregistered benchmarks whose frozen result artifacts — misses
included — live in bench/ (
results/*.json, each carrying
the declared terms it was scored under), and bettermemory eval
scores whether memory actually helped against your own log (ours
is published).
For agents
You were likely handed this link with "set this up." On Claude Code,
the two plugin commands above are the whole install — skip to step 3.
Everywhere else, steps are idempotent and safe to re-run:
uv tool install bettermemory — or pipx install bettermemory,
or pip install bettermemory into a venv. Python 3.11–3.14.bettermemory init --client claude-code — likewise
claude-desktop, cursor, cline, hermes. For any other MCP host, plain
bettermemory init prints the canonical mcpServers JSON snippet
plus known config locations. (Continue needs a hand-written YAML
entry — see docs/clients.md.)bettermemory doctor — exit 0 means correctly wired; every failed
check prints a one-line fix.- Have the user restart the client so the server loads, then confirm
by asking the model "what memory tools do you have?"
Your operating contract — tool signatures, retrieval discipline,
write gates — is docs/api.md; the server's
instructions block delivers the core policy automatically. For the
long-form policy in your system prompt: bettermemory init --with-addendum. Migrating from Claude Code's built-in auto-memory:
bettermemory ingest imports those files once.
Everything else
The rest is written down: mechanics, storage format, the full tool
surface, CLI, configuration and limitations in
internals; per-client setup in clients and
installation; eval methodology in eval;
postmortems in incidents; release history in
CHANGELOG; dev setup and the compatibility contract in
CONTRIBUTING.
MIT licensed — see LICENSE.