hermes-context-graph

hermes-context-graph

Hermes Agent binding for Practice D — context graph. A thin ContextEngine over the unchanged agent-context-core; the Card model, scan, scoring, matcher and projection live in context_core.graph.

What it does

Each closed turn becomes a Card. On every request the engine projects the conversation at the resolution this question needs — Full Content, Description, or just Title — so the model sees a compact graph instead of the whole transcript. The persisted history is never deleted (projection, not destruction): a mis-cut costs one recovery call, not a lost fact. Three tools raise folded content back: expand_card(titles), expand_artifact(reference), find_context(need).

How it maps onto the Hermes ContextEngine

Seam	Role
on_turn_complete(messages)	store each tool return as an addressable artifact under <tool_call_id>_0; derive its artifact Cards
select_context(request_messages)	run context_core.graph.project over closed turns; return the projected list; request-only; fail-open
get_tool_schemas() / handle_tool_call()	expand_card + expand_artifact + find_context

The graph state and reference store live on the engine instance (process-local, not Hermes persisted state — the same rationale as the LangGraph binding's per-conversation store).

Single-select bonus

Because this engine replaces Hermes's default lossy context_compressor, the "nothing may delete from the history behind the graph" precondition is satisfied structurally — there is no co-active summarizer to fight (in LangGraph the middleware can only warn about a co-installed pruning middleware; here single-select makes the conflict impossible).

Install & select

pip install hermes-context-graph
pip install "hermes-context-graph[hermes]"  # + Hermes host from source

context:
  engine: context-graph

Pass stash= a relevance filter's store so a [ref: …] the filter minted also resolves through expand_artifact (this is what hermes-all-three wires up).