Persistent memory: how CodeGraph stops your agent from relearning your repo

[anvanster]

Your agent has amnesia. CodeGraph fixes it.

Most MCP setups give the model tools but no memory. Every new session starts cold: it re-greps the same files, re-derives the same call chains, and forgets the decisions you explained last time. CodeGraph persists two layers so that knowledge compounds instead of resetting.

1. Structural memory (automatic).
The parsed graph and embeddings are written to a local RocksDB under ~/.codegraph. Restart the server and it reloads the index - no re-parsing, no re-embedding. The structural understanding of your repo survives across sessions for free.

2. Working memory (explicit).
Three MCP tools let the agent carry insights forward:

• memory_store - save an insight, decision, or gotcha, scoped to a file, module, or the whole project.
• memory_search - semantic recall across everything ever stored, from any later session.
• memory_context - auto-pull the memories relevant to the current file or task.

Both layers are 100% local - nothing is uploaded to remember it, which also makes this work air-gapped.

Try it

"codegraph": {
  "command": "codegraph-server",
  "args": ["--mcp"]
}

Point any MCP client (Claude Code, Cursor, Windsurf, Codex) at it, index once, and the next session already knows your codebase.

Part of 42 MCP tools across 38 languages, in a single Rust binary. Apache-2.0.

How are you using the memory tools? If you have a workflow where cross-session memory saved you real re-priming time, drop it below - we are collecting patterns for the docs.