Memorine

Persistent memory for OpenClaw agents. No APIs, no Docker, no external services. Just Python and SQLite.

What it does

• Facts -- store, recall, and search with full-text search (FTS5). Near-duplicates are reinforced, contradictions are flagged automatically.
• Forgetting curve -- memories decay over time if not accessed, just like human memory. Old garbage cleans itself up.
• Events -- log what happened with causal chains (this caused that, which caused this other thing).
• Procedures -- track multi-step workflows, learn which steps tend to fail, anticipate what you will need before starting a task.
• Cross-agent sharing -- agents share facts with each other or the whole team through the same database.
• Semantic search (optional) -- install memorine[embeddings] to add meaning-based search on top of keyword search.

14 MCP tools

Once installed, your agents get access to:

Setup

Install Memorine:

pip install memorine

Add the MCP server to your OpenClaw config (openclaw.json):

{
  "mcpServers": {
    "memorine": {
      "command": "python3",
      "args": ["-m", "memorine.mcp_server"]
    }
  }
}

Allow the tools for your agents:

{
  "tools": {
    "allow": [
      "memorine_learn",
      "memorine_recall",
      "memorine_log_event",
      "memorine_events",
      "memorine_share",
      "memorine_team_knowledge",
      "memorine_profile",
      "memorine_anticipate",
      "memorine_procedure_start",
      "memorine_procedure_step",
      "memorine_procedure_complete",
      "memorine_correct",
      "memorine_stats",
      "memorine_learn_batch"
    ]
  }
}

Optional semantic search (adds meaning-based matching on top of keyword search):

pip install memorine[embeddings]

Security

Each agent can only modify its own data. Forget, correct, link, and procedure operations all validate ownership. No agent can touch another agent's memories.

How it works

Everything lives in a single SQLite file at ~/.memorine/memorine.db. FTS5 handles keyword search, triggers keep indexes in sync, and WAL mode allows concurrent reads. The forgetting curve uses retention = e^(-days / stability) where stability grows each time a memory is accessed. No LLM needed for any of this.