codemem

Your coding agents keep rebuilding what you already have. codemem remembers what they built, across every project and every machine, and scores each piece on evidence of whether it worked.

It is an MCP server for Claude Code and other MCP clients. It knows what projects exist, where they live, what reusable pieces they contain, what each session did and decided, every commit pushed to the git host, and how to do recurring things. Everything is searchable from any machine, from an agent, or from the web UI. It is written to during work, not after.

One Python process, SQLite, no external services required.

What makes it different

Inventory scripts snapshot a directory and you read the report once. Registries list what people published. codemem does neither.

• Trust from evidence, not from stars. Every project and asset gets a computed score from how recently it changed, how much work went into it, whether it is actually deployed and running, whether anything else imports it, and whether a human vouched for it. verify records "I ran it today and it works" and restarts the freshness clock. Scores are recomputed on every sync, so they decay when you stop touching something.
• It finds the copies you forgot. Discovery mines every repo at HEAD for scripts, modules, service units, Dockerfiles and MCP servers, hashes their functions, and links repos that share code. It will tell you that you wrote the same retry wrapper in three places.
• Agents feed it automatically. A session hook records what each Claude Code session did, and gives the next session a brief on the project it just opened. The catalogue builds itself.
• Maturity is a first-class label. authoritative, usable, experimental, antiquated, sunset, broken, junk, each with a reason. Search ranks by it, and you can exclude the ratings you should not build on.
• Your code stays where it is. codemem stores metadata: names, descriptions, imports, symbols, function hashes, commit records. It does not copy your source anywhere.

Endpoints

Normally runs on the git host, the machine holding the bare repos, so it can read them directly.

Security: codemem has no authentication. It is built for a trusted network, and anyone who can reach the port can read and write everything. Do not expose it to the internet without an authenticating proxy in front of it. See docs/architecture.md.

Requirements

Python 3.11 or newer, and git. Optional: a git host with bare repos for the commit feed, Gitea for descriptions and web links, and Ollama for semantic search, drafted descriptions and the review pass. Everything works without the optional pieces; search falls back to BM25 alone.

Why

A developer with many project folders across several machines has no way for a new Claude Code session to know that a retry wrapper, a Gradio helper, a GPU monitor or a data pipeline already exists somewhere. So they get rebuilt. Snapshot inventory scripts (run, read, forget) do not fix that. codemem is the live, writable layer: it is written to during work, not after.

The content is whatever is there. codemem does not filter or judge. Each project carries an audience: unrestricted (the default), professional, or employer, so a search or a view can exclude one when the context calls for it. Nothing is hidden unless asked, with one exception: cloned third-party repos are marked origin = vendor and stay out of search and lists until you ask for them. See docs/USER_GUIDE.md.

Install from PyPI

The package is codemem-mcp; the command it installs is codemem.

pip install codemem-mcp          # or: uv tool install codemem-mcp / pipx install codemem-mcp
codemem serve                    # MCP at http://localhost:8055/mcp, web UI at http://localhost:8055/

uvx codemem-mcp serve            # or run it without installing

Data goes to ~/.codemem/ unless CODEMEM_DB or CODEMEM_DATA says otherwise. Register it with Claude Code:

claude mcp add --transport http --scope user codemem http://localhost:8055/mcp

That is the server alone. For the systemd units, the commit feed and the session hooks, use the quick start below from a clone of the repository.

Run with Docker

docker build -t codemem .
docker run --rm -p 127.0.0.1:8055:8055 codemem                        # throwaway database
docker run -d -p 127.0.0.1:8055:8055 -v codemem-data:/data codemem    # keep it in a volume

No authentication. Anyone who can reach the port can read and write everything. Keep the 127.0.0.1: in -p: a bare -p 8055:8055 publishes the port on every interface, and Docker's port rules bypass most host firewalls.

The image needs none of the optional pieces: embeddings are off (CODEMEM_EMBED=0) and there is no git host. Pass -e OLLAMA_URL=... -e CODEMEM_EMBED=1 to turn semantic search on.

Quick start

Server (once, on the git host):

./install.sh                # venv, user-scope systemd units, seed, backfill, gitea, scan, embed

Client (each machine, once):

CODEMEM_URL=http://<host>:8055 client/install-client.sh    # Linux/macOS: MCP server + SessionStart/SessionEnd hooks
client\install-client.ps1                                   # Windows PowerShell (set $env:CODEMEM_URL first)

Then in any Claude Code session the codemem tools are available and every session starts with a brief of the current project (if codemem knows it) and ends with an automatic session record.

What Claude sees

Where the data comes from

• Git host: every bare repo in GIT_ROOT is backfilled; a post-receive hook posts each push to /ingest/push so new commits appear within seconds. See docs/git-commit-feed.md.
• Gitea (optional): descriptions, topics and web URLs, via its API with a read token.
• Local scan: project directories under registered roots, per machine. On the server this is automatic; other machines run client/codemem_agent.py once a root is registered.
• Discovery: every own repo is mined at HEAD for scripts, modules, units, Dockerfiles, prompts and MCP servers, recorded as auto-discovered assets, with shared-code links between repos.
• Docs: markdown from CODEMEM_DOC_SOURCES (this repo by default), re-hashed every sync.
• Claude Code: session records from the hook, plus whatever Claude writes with the tools.
• Local model (optional, via Ollama): embeddings for semantic search, drafted descriptions, and a targeted review of near-duplicate functions and thinly described code.

Layout