Hermes Local Memory is a SQLite-based memory provider for Hermes Agent. It keeps profiles, peers, aliases, raw message history, durable facts, compact cards, session summaries, full-text search, and prompt context injection in one local database file (by default ~/.hermes/memory/local_memory.sqlite), with no server, queue, vector store, or daemon.
The design treats memory as something the agent inspects and maintains through explicit tools rather than an opaque side process. Identity is data: aliases such as telegram:1001 or honcho:Alice map to canonical peers, and a scheduled peer-review packet lets the agent map new platform identities or escalate ambiguous ones. Raw history is preserved on import and is not rewritten by identity repair unless a tool says so.
Consolidation is deterministic and conservative. Dry runs produce bounded plans that can promote safe candidate facts, supersede duplicates, and bootstrap empty cards, but imported candidates are never bulk-promoted and compact cards are curated or replaced rather than grown by appending every fact. Imports from Hermes' built-in USER.md and MEMORY.md markdown memory and from the Honcho HTTP API are additive and idempotent, support identity maps, and back up the target database. The project is at alpha status.
Features
- Provider tools: memorycontext, memorysearch (FTS5 over active facts), memoryconclude (facts with evidence links), memorygetcard and memoryset_card
- Maintenance tools: memoryconsolidate per peer and memorymaintenance across all subject and observer pairs, each with preview and apply modes
- Review packets: build-and-apply pairs for peer identity, reflection over stale sessions, candidate facts, compact cards, and first Honcho migration
- CLI inspection: list peers, aliases, sessions, cards, messages, and facts, and render the exact context that would be injected, with JSON output
- Explicit repair: alias add, fact add and retract, and full-card replacement from JSON
- Markdown import: USER.md entries become user facts and cards, MEMORY.md entries become the assistant's self facts and self-card
- Honcho import: dry-run and apply through the Honcho API with identity maps for fragmented peers, never mutating Honcho
- Plugin shim: install-shim registers the provider and a namespaced local_memory:maintenance skill without changing config.yaml or switching the live provider
- Data model: profiles, peers, peeraliases, sessions, sessionpeers, messages, facts, cards, and summaries tables with FTS on messages and facts
