Summary
docs/honcho-integration-spec.md explicitly calls out several openclaw-honcho patterns that Hermes should adopt back: message dedup (lastSavedIndex) and platform metadata stripping before Honcho storage. The current Honcho plugin already has async prefetch, dynamic reasoning, identity seeding, and session naming, but it still appears to write session messages without a dedup cursor and without a dedicated metadata-stripping layer.
Evidence
Docs
docs/honcho-integration-spec.md:92-98
lastSavedIndex dedup
- platform metadata stripping
docs/honcho-integration-spec.md:338-349
- openclaw-honcho checklist lists these as already-done there and not yet re-implemented in Hermes
Current implementation
plugins/memory/honcho/session.py:324-359
_flush_session() gathers all local unsynced messages and sends them via honcho_session.add_messages(...)
- no
lastSavedIndex-style persisted cursor was found
- repo search did not find
lastSavedIndex in the Python implementation
- repo search did not find a Honcho-specific platform metadata stripping layer or equivalent normalizer near the Honcho sync path
Problem
Without a persisted dedup cursor and a metadata hygiene layer:
- replay / duplication risk is higher across restarts, reloads, and rehydration paths
- transport-specific metadata may leak into long-term memory if it enters message content or wrappers upstream
- Hermes Honcho storage is less robust than the documented target architecture
Scope
- Add a persisted dedup mechanism for Honcho session writes
- equivalent to
lastSavedIndex semantics or a stronger cursor-based variant
- Ensure replays / restored local history do not resend already-synced messages
- Add platform metadata stripping / normalization before Honcho persistence
- remove adapter-only transport noise
- keep user/assistant semantic content intact
- Add tests for restart/reload/retry scenarios and dedup correctness
Acceptance criteria
- Repeated flushes cannot duplicate previously synced messages after process restart or cache rebuild
- Honcho persistence path strips platform-only metadata before storage
- Tests cover duplicate-prevention and metadata hygiene
docs/honcho-integration-spec.md can be updated to mark these Hermes backports complete
Related
Summary
docs/honcho-integration-spec.mdexplicitly calls out several openclaw-honcho patterns that Hermes should adopt back: message dedup (lastSavedIndex) and platform metadata stripping before Honcho storage. The current Honcho plugin already has async prefetch, dynamic reasoning, identity seeding, and session naming, but it still appears to write session messages without a dedup cursor and without a dedicated metadata-stripping layer.Evidence
Docs
docs/honcho-integration-spec.md:92-98lastSavedIndexdedupdocs/honcho-integration-spec.md:338-349Current implementation
plugins/memory/honcho/session.py:324-359_flush_session()gathers all local unsynced messages and sends them viahoncho_session.add_messages(...)lastSavedIndex-style persisted cursor was foundlastSavedIndexin the Python implementationProblem
Without a persisted dedup cursor and a metadata hygiene layer:
Scope
lastSavedIndexsemantics or a stronger cursor-based variantAcceptance criteria
docs/honcho-integration-spec.mdcan be updated to mark these Hermes backports completeRelated
docs/honcho-integration-spec.md