nodecode-hydradb

HydraDB memory: recall before you answer, remember a note, forget one, and capture every turn. A Nodecode rewrite of hydra-db/hydradb-hermes-plugin by hydra-db.

The model gets three tools — hydradb_recall, hydradb_remember, hydradb_forget — over HydraDB's long-term memory, /hydradb to show the backend or recall from the shell, and a worker that stores each finished turn in the background.

Get it

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

Set it up

The cell needs a HydraDB account: an API key and the database (tenant) it reads and writes. Either export them, as the original does —

export HYDRADB_API_KEY="…"
export HYDRADB_DATABASE="my-database"

— or write the hydradb section of config.jsonc (or use /settings):

{
  "hydradb": {
    "api_key": "…",               // or api_key_env; HYDRADB_API_KEY is read either way
    "database": "my-database",    // or HYDRADB_DATABASE
    "collection": "my-collection" // optional sub-tenant, or HYDRADB_COLLECTION
  }
}

A start with no key and no database is refused, naming the member.

Member	Default	What it does
`api_key`	unset	the HydraDB API key; `api_key_env` and `HYDRADB_API_KEY` are read too
`database`	`HYDRADB_DATABASE`	the database (tenant) recall reads and capture writes to
`collection`	`HYDRADB_COLLECTION`	the sub-tenant inside the database
`base_url`	`https://api.hydradb.com`	the API address; `HYDRADB_BASE_URL` is read too
`kind`	`knowledge`	the family recall searches, as the original recalls
`search_limit`	`6`	how many passages one recall asks for unless the call says otherwise
`infer`	`true`	let HydraDB extract facts from what `hydradb_remember` stores
`infer_capture`	`false`	let it extract facts from a captured turn too, as the original does not
`capture`	`true`	store each finished turn in the background
`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.

The model calls hydradb_remember. Later, "what do I drink?" makes it call hydradb_recall, and HydraDB answers with the passages nearest the question and their ids — the ids hydradb_forget takes. /hydradb prints the database, the collection, the address and whether capture is on; /hydradb where do I live recalls from the shell.

With capture on (the default), every finished turn goes to HydraDB in the background as one conversation note, User: … then Assistant: …, exactly as the original's sync_turn writes it — no extraction unless infer_capture asks for one. Five consecutive backend failures pause the capture path for two minutes, and the eight newest turns are the ones queued. (help :hydradb) is the page the model reads.

The three endpoints

HydraDB's public API, and nothing else:

POST /query — hybrid recall, JSON body {tenant_id, sub_tenant_id, query, kind, mode, max_results}. The answer's llm_prompt is used as-is when the server renders one; else the chunks become the lines the model reads, each chunk's stored content envelope unwrapped to its prose and any trailing graph enrichment kept.
POST /context/ingest — multipart, app_knowledge a JSON array of {id, title, content: {text}, metadata, infer}; ingestion is asynchronous.
DELETE /context — JSON body {tenant_id, sub_tenant_id, ids, kind}.

All three authenticate with Authorization: Bearer <key>.

Not ported

Recall before a reply. The original's prefetch injects what HydraDB finds into the prompt of every turn. Nodecode's standing-section seam reads text once per turn and carries no per-message query, so the model calls hydradb_recall (the tool's description and (help :hydradb) say when); the capture after a turn is still automatic.
hermes memory setup / hermes memory status. The section, /settings and /hydradb are the same surface.
The MCP server the repository mentions. hydradb_recall is the pull-style recall it would have given, without another process.

Apache-2.0, HydraDB's.