bettermemory
<p align="center">
<img src="https://raw.githubusercontent.com/0Mattias/bettermemory/main/docs/assets/banner.svg" alt="bettermemory: memory that is checked before it is believed" width="100%">
</p>
<p align="center">
<a href="https://pypi.org/project/bettermemory/"><img src="https://img.shields.io/pypi/v/bettermemory.svg" alt="PyPI"></a>
<a href="https://github.com/0Mattias/bettermemory/actions/workflows/ci.yml"><img src="https://github.com/0Mattias/bettermemory/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
<a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.11%E2%80%933.14-blue.svg" alt="Python 3.11-3.14"></a>
<a href="LICENSE"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="MIT"></a>
<a href="https://glama.ai/mcp/servers/0Mattias/bettermemory"><img src="https://glama.ai/mcp/servers/0Mattias/bettermemory/badges/score.svg" alt="bettermemory MCP server"></a>
</p>
<!-- mcp-name: io.github.0Mattias/bettermemory -->
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:
```sh
uv tool install bettermemory
bettermemory init --client claude-desktop # or cursor / cline / claude-code / hermes
```
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/][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][eval-results]).
## 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:
1. `uv tool install bettermemory` — or `pipx install bettermemory`,
or `pip install bettermemory` into a venv. Python 3.11–3.14.
2. `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][clients-continue].)
3. `bettermemory doctor` — exit 0 means correctly wired; every failed
check prints a one-line fix.
4. 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][api]; 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][internals]; per-client setup in [clients][clients] and
[installation][installation]; eval methodology in [eval][eval];
postmortems in [incidents][incidents]; release history in
[CHANGELOG][changelog]; dev setup and the compatibility contract in
[CONTRIBUTING][contributing].
MIT licensed — see [LICENSE][license].
[api]: https://github.com/0Mattias/bettermemory/blob/main/docs/api.md
[bench]: https://github.com/0Mattias/bettermemory/tree/main/bench
[changelog]: https://github.com/0Mattias/bettermemory/blob/main/CHANGELOG.md
[clients]: https://github.com/0Mattias/bettermemory/blob/main/docs/clients.md
[clients-continue]: https://github.com/0Mattias/bettermemory/blob/main/docs/clients.md#continue-legacy-shape--see-caveat
[contributing]: https://github.com/0Mattias/bettermemory/blob/main/CONTRIBUTING.md
[eval]: https://github.com/0Mattias/bettermemory/blob/main/docs/eval.md
[eval-results]: https://github.com/0Mattias/bettermemory/blob/main/docs/eval-results.md
[incidents]: https://github.com/0Mattias/bettermemory/blob/main/docs/incidents/
[installation]: https://github.com/0Mattias/bettermemory/blob/main/docs/installation.md
[internals]: https://github.com/0Mattias/bettermemory/blob/main/docs/internals.md
[license]: https://github.com/0Mattias/bettermemory/blob/main/LICENSE
TDQS
Scored across 18 tools
The memory_* and episode_* prefixes cleanly separate two tiers (durable memories vs. journal entries), and within each tier the operation verbs (search/list/show vs. write/update/remove/verify) are clearly distinct. The confirm/cancel and scope_disable/enable pairs are unambiguous complements. Even the two search tools are well-differentiated: memory_search is ranked while episode_search is explicitly chronological.
All 18 tools follow a strict resource_prefix + action_verb pattern in snake_case: memory_search, memory_write, episode_handoff, memory_scope_disable. Compound names like memory_write_confirm and memory_scope_overview extend the pattern predictably. No camelCase, no vague verbs, no mixed conventions.
18 tools is slightly above the ideal 3-15 band, but the domain is genuinely rich — memory lifecycle, verification, drift detection, scoping, episodes, and audit all earn dedicated tools. One or two (memory_audit_turn, which is hook-dispatched) could arguably be trimmed from the model-facing surface, but none feels redundant.
The core CRUD lifecycle is well covered (write/update/remove/show/verify/search), but the tool descriptions repeatedly reference tools that are absent from the exposed surface: memory_restore, memory_list_tombstones, memory_health, memory_proposals, and memory_conflicts are all invoked as if callable but not listed. An agent following the memory_remove description's 'use memory_restore(id) to undo' instruction would hit a dead end. There is also no episode_update or episode_remove, leaving the journal tier with no correction path other than promotion.