Spectron ⇄ Hermes Agent

A Hermes Agent memory provider backed by SurrealDB Spectron — provenance-first, tri-temporal agent memory with semantic, lexical, graph and temporal recall.

Once installed and selected, the agent automatically:

• recalls relevant memories before every turn,
• writes each completed turn back to Spectron (asynchronously),
• consolidates memory when a session ends,

and gains six explicit memory tools it can call directly.

Requirements

• Python 3.10+
• A running Hermes Agent install
• Spectron access (endpoint, context, API key). Spectron is in invite-only preview.

Install

pip install spectron-hermes

This pulls in the SurrealDB SDK (surrealdb v3, which bundles Spectron) and registers the plugin with Hermes via the hermes_agent.plugins entry point.

Alternative: drop-in directory

If your Hermes setup discovers memory providers by directory, copy the package into $HERMES_HOME/plugins/spectron/ (it ships a plugin.yaml manifest for this) and pip install "surrealdb>=3.0.0a1".

Configure & activate

Provide credentials via the environment (the API key is a secret and belongs in .env):

export SPECTRON_API_KEY="..."
export SPECTRON_ENDPOINT="https://your-instance.spectron.dev"
export SPECTRON_CONTEXT="my-context"

Then select the provider — Hermes will prompt for any missing settings:

hermes memory setup      # choose "spectron"
hermes memory status     # confirm it is active
hermes                   # run a session with Spectron-backed memory

Non-secret settings are written to $HERMES_HOME/spectron.json.

Configuration

Setting	Env var	Default	Notes
api_key	SPECTRON_API_KEY	—	secret, required (stored in .env)
endpoint	SPECTRON_ENDPOINT	—	required, origin with no trailing slash
context	SPECTRON_CONTEXT	—	required; Spectron pins a client to one context
recall_mode	SPECTRON_RECALL_MODE	hybrid	hybrid | context | tools
write_frequency	SPECTRON_WRITE_FREQUENCY	turn	turn | session
top_k	SPECTRON_TOP_K	5	memories recalled per turn
default_scope	SPECTRON_DEFAULT_SCOPE	—	e.g. user/tobie; scope for writes / lens for reads

Recall modes: hybrid injects raw recalled memories before each turn; context injects a synthesised answer instead; tools injects nothing and lets the model recall explicitly via spectron_recall / spectron_context.

Tools exposed to the agent

Tool	Spectron call	Purpose
spectron_recall(query, k?)	recall	Search memory (semantic/lexical/graph/temporal).
spectron_remember(text, scope?)	remember	Store a durable fact.
spectron_context(query, k?)	query_context	Synthesised answer from memory.
spectron_forget(query, purge?)	forget	Supersede (default) or hard-delete.
spectron_reflect(query, persist?)	reflect	Derive insights; optionally persist.
spectron_upload(path, title?)	documents.upload	Ingest a document into knowledge memory.

Reliability

The provider is built to never destabilise the agent:

• Writes run on a background daemon thread — turns never block on I/O.
• Every Spectron call is wrapped; failures are logged and degrade to empty results rather than raising into the agent loop (fail open).
• After repeated failures (or an auth error) a circuit breaker disables memory for the rest of the session.

Development

pip install -e ".[dev]"
pytest

Tests use a mock Spectron client and need neither a live server nor the surrealdb SDK installed.

License

Apache-2.0