Align Hermes passive memory recall with OpenClaw six-layer DKG memory behavior

[Jurij89]

Context

During Hermes adapter E2E testing on PR #315, we confirmed that the Hermes adapter tools are registering and core DKG tool flows are working, including status, context graph list/create, assertion create/write/query/promote, SWM query, sub-graph create/list, and memory_search.

One follow-up gap remains around passive memory behavior: Hermes currently does not fully mirror the latest OpenClaw DKG memory recall architecture.

Current Behavior

OpenClaw passive recall now fans out across six DKG memory layers:

• agent-context / working-memory
• agent-context / shared-working-memory
• agent-context / verified-memory
• selected project/context graph / working-memory
• selected project/context graph / shared-working-memory
• selected project/context graph / verified-memory

Chat turns are no longer the only passive recall source. Chat turns are included because they land in agent-context working memory, but OpenClaw no longer pins recall to only the chat-turns assertion.

Hermes currently has split behavior:

• memory_search already performs the expected six-layer fan-out.
• Passive Hermes memory prefetch() still reads only the configured provider memory assertion, currently defaulting to agent-context / memory.
• Passive injected memory context is not labeled with enough provenance for the agent to reliably know which context graph and memory layer each recalled item came from.
• dkg_memory writes to the configured provider memory assertion only. It does not currently accept an optional context_graph_id for project-scoped memory writes.

Desired Behavior

Hermes passive memory recall should mirror OpenClaw and the SKILL.md contract:

1. Passive recall should always query agent-context across WM/SWM/VM.
2. If a project/context graph is selected or otherwise supplied by the calling surface, passive recall should also query that graph across WM/SWM/VM.
3. Passive recall should inject recalled snippets with explicit provenance so the agent can distinguish:
   • context graph id
   • memory layer/view
   • score/source where available
4. The injected context should remain safe for chat persistence, i.e. avoid recalled-memory boomerang into saved chat turns.
5. Consider extending dkg_memory with optional context_graph_id, defaulting to agent-context, so agents can store simple memory notes in a specific project/context graph when the user asks for that explicitly.

Suggested Implementation Notes

• Reuse or share the same fan-out logic already used by Hermes memory_search rather than creating a second query planner.
• Match OpenClaw layer labels where possible:
  • agent-context-wm
  • agent-context-swm
  • agent-context-vm
  • project-wm
  • project-swm
  • project-vm
• Include the exact context_graph_id in Hermes injected passive context, even where OpenClaw currently only implies it through project-* labels.
• Keep the provider default simple: agent-context / memory remains the default write target for dkg_memory unless a specific context graph is supplied.
• Preserve existing assertion tools as the precise interface for explicit WM/SWM workflows.

Acceptance Criteria

• Hermes passive memory prefetch() queries all three memory layers for agent-context.
• Hermes passive memory prefetch() queries all three memory layers for the selected project/context graph when available.
• Passive injected memory entries include enough provenance for the agent to know the context graph and memory layer.
• memory_search and passive recall behavior are aligned and covered by tests.
• If dkg_memory gains optional context_graph_id, the schema, handler, docs, and tests are updated.
• Regression tests cover:
  • agent-context WM/SWM/VM passive recall
  • selected project WM/SWM/VM passive recall
  • no selected project fallback to agent-context only
  • provenance labels in injected context
  • no recalled-context boomerang into persisted chat turns, if applicable

Related Work

• Follow-up from Hermes adapter PR Integrate Hermes adapter with DKG local agent #315.
• Mirrors latest OpenClaw passive DKG memory behavior.
• Related skill contract: packages/cli/skills/dkg-node/SKILL.md memory recall / memory_search behavior.