A native Hermes Agent memory provider backed by a local EverOS server: markdown files as the source of truth, hybrid (vector + BM25 + scalar) recall, and two memory tracks — your profile/episodes and the agent's distilled cases/skills. Recall is automatic every turn; the model gets zero memory tools to remember to press.
Design details: docs/DESIGN_DOC.md · build/test/release:
docs/HANDOFF.md.
hermes plugins install EverMind-AI/plugins/hermes # → ~/.hermes/plugins/memory/everos/
hermes memory setup # pick "everos"; answer the prompts
hermes memory status # verify: provider active, EverOS healthyOne-time, outside the plugin: EverOS needs everos init once and API keys in
~/.everos/everos.toml. If EverOS isn't running, the provider starts it
automatically on agent startup (configure everos_dir / start_cmd).
Updating: subfolder installs can't use hermes plugins update — re-run the
install command above.
| Key | Default | Purpose |
|---|---|---|
base_url |
http://127.0.0.1:8000 |
Where EverOS listens (its port is EverOS config) |
user_id |
OS account | Developer identity for the user track |
agent_id |
hermes |
Constant pooled agent identity |
query_max_units |
500 |
Weighted recall-query clip (CJK char = 2 units) |
everos_dir |
— | EverOS checkout, for auto-start from its venv |
start_cmd |
everos server start |
Auto-start command (quote spaced paths) |
The provider talks only to your local EverOS (127.0.0.1). EverOS itself sends
conversation text (each completed turn, including tool calls/results; built-in
memory edits; delegation summaries) to the LLM and embedding providers you
configured in EverOS for extraction — cloud services unless you point EverOS
at local models. Nothing else leaves the machine; the memory store
(Markdown/SQLite/LanceDB) is local.