anamnesis-mcp

MCP stdio server (and one-shot CLI) for anamnesis — associative cognitive memory for LLM agents.

Tools

| Tool | When the agent should call it | |------|-------------------------------| | recall | Before answering — surfaces prior decisions/lessons. Reading auto-reinforces what it returns. | | remember | After any decision or lesson worth keeping — stores one distilled insight. | | ingest_conversation | Hand off a full transcript (ordered turns) for the windowing recipe. |

Install (Claude Desktop)

// claude_desktop_config.json
{
  "mcpServers": {
    "anamnesis": {
      "command": "npx",
      "args": ["-y", "anamnesis-mcp", "serve"],
      "env": { "ANAMNESIS_DB": "/Users/you/.anamnesis/memory.db" }
    }
  }
}

First run downloads the embedding model (~400 MB) to a per-user cache. Pre-warm interactively with npx anamnesis-mcp prewarm.

The npm package is a small launcher. During install it downloads the matching native anamnesis-mcp binary from the GitHub Release for the package version. Set ANAMNESIS_MCP_BINARY to use a locally built binary instead, or ANAMNESIS_MCP_SKIP_DOWNLOAD=1 when packaging without downloading.

Make memory reliable (the important part)

This is a general MCP server — there are no hooks, so the agent must choose to call the tools. Paste this into your system/project instructions so it does:

Before answering anything non-trivial, call recall with the key terms of the question. After you decide something, reach a conclusion, or learn a lesson the user would want kept, call remember with a one-sentence statement of it.

Configuration

| Env var | Default | Meaning | |---------|---------|---------| | ANAMNESIS_DB | <data_dir>/anamnesis/memory.db | SQLite file for the default namespace. | | ANAMNESIS_NAMESPACE | default | Namespace when a call omits one. | | ANAMNESIS_REINFORCE | true | Auto-commit (reinforce) recalled results. Set false for receipt mode. | | FASTEMBED_CACHE_DIR | <cache_dir>/anamnesis/models | Where the bge model is cached. |

Namespaces are isolated by separate SQLite files (<db_dir>/<namespace>.db).