nodecode-supermemory

Supermemory: long-term memory that survives sessions, with the four tools the model calls to save, search, forget and profile it. A Nodecode rewrite of supermemoryai/hermes-supermemory by supermemoryai.

The model gets four tools — supermemory_store, supermemory_search, supermemory_forget, supermemory_profile — that keep a user's memories in one Supermemory container and find them again by meaning, /supermemory to show the container or search it, and a worker that writes each finished turn to the container as a conversation, so facts are extracted without the model being asked.

Get it

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

Set it up

The cell needs a Supermemory API key — a key is made at https://console.supermemory.ai/keys?create=true — and the supermemory section of config.jsonc (or /settings) names it:

{
  "supermemory": {
    "api_key_env": "SUPERMEMORY_API_KEY",   // or "api_key" inline
    "container_tag": "nodecode"
  }
}

SUPERMEMORY_API_KEY in the environment is read either way. A missing key is refused at the start, naming the member; a wrong search_mode, an integer under its floor and a bad base_url are refused the same way.

Member	Default	What it does
`api_key`	unset	the Supermemory API key
`api_key_env`	unset	the environment variable holding it (`SUPERMEMORY_API_KEY` is read too)
`base_url`	`https://api.supermemory.ai`	the API address; a self-hosted server answering the same shapes goes here
`container_tag`	`nodecode`	the data partition: the same tag across sessions finds the same memories
`search_mode`	`hybrid`	`hybrid` (memories with their source chunks), `memories` or `documents`
`search_limit`	`10`	how many memories one search asks for unless the call says otherwise
`capture`	`true`	write each finished turn to the container
`entity_context`	the built-in	what a capture is worth extracting, at most 1500 characters
`timeout_seconds`	`20`	one request's ceiling

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 supermemory_store. Later, "what do I drink?" makes it call supermemory_search first, and Supermemory answers with the memory and its id — the id supermemory_forget takes.

/supermemory prints the container, the search mode and whether capture is on; /supermemory where do I live searches memory from the shell.

With capture on (the default), every finished turn is queued in the background and written as a document in the [role: user]…[user:end] layout of the original, one document per session per 4-hour window, so the API appends a session's turns to one document instead of starting another. The entity context tells Supermemory what in a conversation is worth remembering. A batch the service refused stays queued, and the queue keeps the newest 50 turns (256 KiB) so a service that is down cannot grow it forever.

The wire

POST /v3/documents with content, containerTags, metadata (stamped sm_source: nodecode), customId and entityContext.
POST /v3/search with q, containerTag, limit and searchMode.
POST /v4/profile with containerTag and, when the tool got one, q.
DELETE /v4/memories with id and containerTag.

Every request carries Authorization: Bearer <key> and x-sm-source: nodecode. The source stamp is the original's routing key: in the Supermemory app it puts a write in that source agent's Space, so a user can filter and bulk-manage what this cell wrote.

Not ported

Recall before a reply. The original searches on the turn's first message and injects a <supermemory-context> block before the model answers. Nodecode's recall seam reads standing sections once per turn and carries no per-message query, so here the model calls supermemory_search (the tools' descriptions and (help :supermemory) say when), and only capture is automatic.
The system-prompt block the original prepends, and with it the profile_frequency throttle on the static/dynamic profile split; the supermemory_profile tool and the help line carry the same guidance.
The kebab-case tool aliases (supermemory-save and friends). Nodecode's catalog has one name per tool, and the four snake_case names are the original's own.
Custom containers. The original whitelists extra container tags that a tool call may name; here one section names the container the cell reads and writes, and {identity} templates are not expanded.
hermes memory setup / hermes memory status. The section, /settings and /supermemory are the same surface.