mneme Quickstart¶

Prerequisites¶

Python 3.11+
Ollama running locally
Claude Code installed (or any agent harness with prompt hooks)

Install¶

pip install git+https://github.com/Luizhcrs/mneme.git
ollama pull nomic-embed-text
mneme init
mneme reindex

mneme init creates ~/.claude/mneme/capabilities.yaml with 10 starter capability cards. mneme reindex embeds them into a local SQLite + sqlite-vec store at ~/.claude/mneme/semantic.sqlite.

Wire the hooks¶

Add to ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": "python -m mneme.hooks.user_prompt_submit",
    "PostToolUse": "python -m mneme.hooks.post_tool_use"
  }
}

The UserPromptSubmit hook injects relevant capabilities into every prompt. The PostToolUse hook records successful tool sequences as procedural memory so similar future tasks reuse them.

Verify¶

mneme search "screenshot the homepage"
mneme stats

You should see Playwright capabilities surface for the screenshot query. Open Claude Code, type a prompt that mentions a tool you have installed, and the agent should now consistently pick the correct capability.

Add your own capability cards¶

Edit ~/.claude/mneme/capabilities.yaml. Each entry follows the schema in capability-card-format.md. After editing:

mneme reindex

Offline fallback¶

If Ollama is unavailable but you still want degraded recall, set:

export MNEME_FALLBACK=1

The hook then matches the prompt against triggers substrings (no embeddings).