nc://cognee

Cognee memory: recall, remember and forget a knowledge graph on a local or remote Cognee server. A Nodecode rewrite of topoteretes/cognee-integrations by topoteretes.

The model gets four tools — cognee_recall, cognee_remember, cognee_forget, cognee_switch_dataset — that keep what a user stores in a Cognee knowledge graph and find it again by meaning, and /cognee to show the server or recall from the shell.

Get it

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

Set it up

The cell talks to a cognee server: the one its own package starts on 127.0.0.1:8011, or a remote deployment. That address is the cognee section's base_url. A server on this machine needs no key; a remote one needs one, in COGNEE_API_KEY or the section:

{
  "cognee": {
    "base_url": "https://cognee.example.com",
    "api_key_env": "COGNEE_API_KEY",   // or "api_key" inline
    "dataset": "nodecode"              // whose memories these are
  }
}

A remote base_url with no key is refused at the start, naming the member, as is a value under its floor (top_k, the two timeouts).

Use it

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

Remember that the release branch is cut on Fridays.

The model calls cognee_remember, which stores one document in the dataset and lets Cognee's pipeline cognify it server-side. Later, "when do we cut the release branch?" makes it call cognee_recall, which answers the passages nearest the question. /cognee prints the server and the dataset; /cognee when do we cut the release recalls from the shell.

Forgetting is two-phase, as in the original: cognee_forget with action="find" and terms lists candidate documents with previews of their stored text, the model shows them to the user, and action="forget" with the confirmed data_ids and confirm=true deletes exactly those. Clearing a whole dataset needs everything_in_dataset and an explicit user request.

cognee_switch_dataset lists the datasets the key sees, reports the current one, moves this conversation to another, or returns it to the configured default.

The wire

POST /api/v1/recall (JSON: query, top_k, datasets, scope ["graph"], search_type), POST /api/v1/remember (multipart: datasetName, run_in_background, and the text as the data part, named after its content so a repeat write is the server's no-op rather than a 409), POST /api/v1/forget (dataset, or datasetId + dataId), GET /api/v1/datasets and GET /api/v1/datasets/<id>/data[/<id>/raw]. Authenticated with X-Api-Key whenever a key is set, as the original sends it.

Not ported

• Local-server mode. The original can download cognee==1.6.5 from PyPI into a private venv and start a server for you. A cell bundles no Python: point base_url at a cognee you started, or at a remote one.
• Per-turn capture and the session bridge. The original sends every turn to a session cache and runs improve() at session end to promote it into the graph; here the model calls cognee_remember (its description says when), and nothing is captured behind the user's back.
• cognee_code_search and cognee_recap. The deterministic code-graph lane needs repositories indexed by the original's CLI, and recaps read the harness's own session table; neither has a Nodecode seam here, so their subjects — indexed code, recorded sessions — are not claimed.
• The > Cognee: credit line and the memory-hit counter. Both are Hermes prompt decorations; a cell contributes tools, a command and a manual page, not prompt text.

Apache-2.0, topoteretes's.