Align Hermes passive DKG memory recall

[Jurij89]

Source issue

Closes #385

Summary

• Rebased onto latest main (200d987b) after additional PRs merged, preserving main's newer Hermes setup parity, dkg_share, context-graph privacy, and build behavior.
• Replaced Hermes passive prefetch() assertion-only recall with the same six-layer DKG fan-out helper used by memory_search.
• Added provenance-rich passive recall blocks with context graph id, DKG view, OpenClaw-style layer label, source, score, and path.
• Added scoped dkg_memory.context_graph_id support while keeping unscoped simple memory writes defaulted to agent-context / memory.
• Added provider-side and daemon-side stripping for adapter-generated recalled-memory blocks before Hermes chat persistence/import.
• Addressed review findings for legacy configured-memory fallback, malformed recall-tag preservation, and short-literal recall.

OpenClaw parity notes

• Hermes passive recall always queries agent-context WM/SWM/VM.
• Hermes passive recall also queries project WM/SWM/VM only when the calling surface supplies a selected project via context_graph_id / target_context_graph; static context_graph config does not count as current selection.
• Layer labels match OpenClaw: agent-context-wm, agent-context-swm, agent-context-vm, project-wm, project-swm, project-vm.
• Trust ranking matches OpenClaw weighting: WM x1.0, SWM x1.15, VM x1.3.
• Recall results dedupe by context graph plus memory URI/text hash, keeping the higher-trust layer for the same memory.
• No adapter-to-adapter imports were added; OpenClaw remains the parity reference.

Files changed

• packages/adapter-hermes/hermes-plugin/__init__.py
• packages/adapter-hermes/test/hermes-adapter.test.ts
• packages/cli/src/daemon/routes/hermes.ts
• packages/cli/test/daemon-hermes.test.ts
• packages/adapter-hermes/README.md
• docs/setup/SETUP_HERMES.md
• packages/cli/skills/dkg-node/SKILL.md

Behavioral details

• Passive prefetch uses a shared Hermes-local helper with memory_search for query planning, fan-out, ranking, dedupe, and provenance.
• agent-context recall always runs across WM/SWM/VM. A selected project graph is included only when supplied through context_graph_id / target_context_graph on the call.
• Passive injected context is emitted as <recalled-memory data-source="dkg-auto-recall"> with escaped snippet text and untrusted-context framing.
• dkg_memory.context_graph_id stores scoped notes under separate cache buckets and queues writes with the target graph id. Unscoped writes continue to target agent-context / memory.
• Existing Hermes profiles that previously stored provider memory in a custom configured context_graph are read as a legacy fallback when the new agent-context assertion is empty or unavailable.
• memory_search no longer filters out short literals with a hard STRLEN >= 20 cutoff.
• Malformed recalled-memory openings without a closing tag are preserved instead of truncating persisted assistant text.

Test evidence

Latest rebase sanity, after updating onto main at 200d987b:

• python -m py_compile packages/adapter-hermes/hermes-plugin/__init__.py packages/adapter-hermes/hermes-plugin/client.py passed.
• git diff --check upstream/main...HEAD passed.
• pnpm --filter @origintrail-official/dkg-adapter-hermes test -- hermes-adapter.test.ts passed: 86 tests.
• pnpm --filter @origintrail-official/dkg-core build passed; this refreshed workspace exports after pulling main.
• pnpm --filter @origintrail-official/dkg build passed.
• pnpm --filter @origintrail-official/dkg test -- test/daemon-hermes.test.ts --reporter=verbose passed: 66 tests. The command still prints the known post-run Windows message, The system cannot find the path specified., after exiting successfully.

Earlier parity/review-round evidence retained from the prior rebase:

• pnpm --filter @origintrail-official/dkg-adapter-openclaw test -- before-prompt-build-hook.test.ts memory-integration.test.ts memory-search-tool.test.ts ChatTurnWriter.test.ts ChatTurnWriter.sanitize.test.ts plugin.test.ts dkg-client.test.ts passed: 7 files, 451 tests.
• pnpm --filter @origintrail-official/dkg-agent... build passed.
• pnpm --filter @origintrail-official/dkg-epcis build passed.

Manual verification status

• No live Hermes gateway or Node UI E2E was run in this workspace.
• Targeted Python provider, daemon route, OpenClaw parity, dependency-build, and CLI build coverage passed locally.
• Hermes provider and QA/security teammate read-only sanity passes found no blockers after the latest rebase.

Security / persistence notes

• Recalled-memory blocks are treated as read-only prompt context and are stripped before provider sync_turn() persistence.
• /api/hermes-channel/persist-turn also strips adapter-generated recalled-memory blocks before chat storage, transition recording, and extraction import.
• Strip tests cover mixed-case sentinel tags, single-quoted data-source attributes, and malformed unmatched openings so recalled context does not boomerang or cause data loss.
• Scoped dkg_memory.context_graph_id avoids cache and queue bleed between default personal memory and project notes.
• Legacy custom-graph provider memory fallback is read-only for recall; new unscoped dkg_memory writes still use agent-context unless context_graph_id is explicit.

Known risks or follow-ups

• No broad refactor into a neutral cross-adapter recall module was attempted in this focused PR.

[github-actions]

🔴 Bug: This early return makes the legacy fallback stop working as soon as agent-context has any note. For upgraded profiles that already have provider memory in a custom context_graph, the first new default write will hide all older notes even though this PR documents that they should remain visible. Merge/dedupe the legacy facts with the new agent-context facts (or migrate them) instead of returning here, and add a regression test for the mixed-data case.

[github-actions]

🟡 Issue: agent-context hits are all tagged as source: "sessions", but dkg_memory also stores durable provider notes in agent-context via urn:hermes:content. That makes the new provenance fields inaccurate and prevents callers from distinguishing simple memory notes from chat-session recall. Derive source from the predicate/assertion type here so Hermes memory entries still report memory.

[github-actions]

🔴 Bug: This only removes well-formed recalled-memory pairs. If the model output is truncated after the sentinel open tag, the recalled snippets survive persistence and get re-indexed on later turns, which is exactly the contamination this change is trying to prevent. Strip sentinelOpen...$ as well here (and keep the Python helper aligned).

[github-actions]

🔴 Bug: context_graph_id now controls where dkg_memory writes, but a malformed non-string value falls through _first_text(...) and silently writes to agent-context. That turns a bad tool call into a data-routing bug. Validate context_graph_id when the key is present and return tool_error on non-string input, like the other DKG handlers do.

[github-actions]

🔴 Bug: Returning here means legacy notes disappear as soon as agent-context contains even one new fact. Upgraded profiles with historical memory in a custom context_graph will lose all of that older context after the first unscoped write. Either merge legacy + default facts during recall or migrate the legacy assertion forward instead of treating it as an all-or-nothing fallback.

[github-actions]

🔴 Bug: context_graph_id is read through _first_text, so any non-string value is silently treated as "missing" and the write falls back to agent-context. A malformed tool call like { "context_graph_id": ["project-cg"] } would store the note in personal memory with no error. Please validate that overrides are non-empty strings before defaulting, and apply the same guard in memory_search, which uses the same parsing pattern.

[github-actions]

🔴 Bug: this legacy read now runs unconditionally for every non-default self._context_graph, but the docs and surrounding comments describe it as a fallback only when agent-context is empty. As written, any future dkg_memory(..., context_graph_id=self._context_graph) note will be merged back into _recall_facts() and leak into the default provider memory prompt. Gate the legacy query behind an empty-default-memory check (or a one-time migration flag) instead of always merging it.