nc://mem0

Mem0 memory: semantic recall and durable facts, through Mem0's cloud or a self-hosted server. A Nodecode rewrite of mem0ai/mem0 by mem0ai.

The model gets four tools — mem0_search, mem0_add, mem0_update, mem0_delete — that keep a user's facts on a Mem0 backend and find them again by meaning, /mem0 to show the backend or search it, and a worker that sends each finished turn to Mem0 for extraction.

Get it

Tell the agent "add the mem0 cell", or run nodecode add nc://mem0.

Set it up

The cell needs a backend, and the mem0 section of config.jsonc (or /settings) names it. Platform mode, Mem0's cloud, needs a key:

{
  "mem0": {
    "mode": "platform",
    "api_key": "…",              // or api_key_env, or just MEM0_API_KEY in the environment
    "user_id": "nodecode-user"   // whose memories these are, the same across sessions
  }
}

Self-hosted mode talks to your own Mem0 server (its FastAPI server/), and needs its address; the key only when the server asks for one:

{
  "mem0": {
    "mode": "self-hosted",
    "base_url": "http://127.0.0.1:8000",
    "api_key_env": "MEM0_SERVER_KEY",
    "user_id": "nodecode-user"
  }
}

A missing key in platform mode is refused at the start, naming the member; a wrong mode or an integer under its floor is refused the same way.

Use it

Ask the model to remember something, then ask about it in a later session:

Remember that I drink green tea and live in Oslo.

The model calls mem0_add. Later, "what do I drink?" makes it call mem0_search first, and Mem0 answers with the fact and its id — the id mem0_update and mem0_delete take. /mem0 prints the backend, its address, the user and the agent; /mem0 where do I live searches it from the shell.

With capture on (the default), every finished turn goes to Mem0 in the background, so facts are extracted without the model being asked; the turn's two messages are cut at a sentence end within max_chars, as the original does for a small embedder's window. A turn whose sync is still busy is skipped rather than queued forever, and five consecutive backend failures pause the capture path for two minutes.

The two wires

• platform: POST /v3/memories/add/, POST /v3/memories/search/, GET/PUT/DELETE /v1/memories/<id>/, authenticated Authorization: Token <key>.
• self-hosted: POST /memories, POST /search, GET/PUT/DELETE /memories/<id>, authenticated X-API-Key when a key is set.

Both are the shapes the original's PlatformBackend and SelfHostedBackend speak, and search is always scoped to filters.user_id so recall spans agents and sessions for one principal, with agent_id and a metadata.channel on every write.

Not ported

• OSS mode. The in-process SDK — a Qdrant or PGVector store, an OpenAI or Ollama model and embedder — is a Python library, and a cell bundles no Python. Platform and self-hosted are the two wires carried.
• Automatic recall before a reply. The original searches on the turn's first message and injects what it finds before the model answers. Here the model calls mem0_search (the tools' descriptions and (help :mem0) say when), and only extraction is automatic. The recall seam Nodecode has reads standing sections once per turn and carries no per-message query.
• hermes memory setup / hermes memory status. The section, /settings and /mem0 are the same surface.
• The PostHog usage events the Python client sends; a request here is the request.

Apache-2.0, mem0ai's.