Skip to main content

Memory systems, the five stores

entity · created 2026-07-09 · actor claude-cowork · confidence high · status active tags: memory kg knowledgelayer memwal

Now​

Layer model (INDEX.md): kg = events, knowledgelayer = distilled truth, instruction layer = behaviour. Concrete stores (structure observed 2026-07-09; store 1 corrected 2026-10-04):

  1. Event bus, the live append-only event store is ~/.hermes/ops/state/events.jsonl, written by ~/.hermes/scripts/hermes_event_bus.py (EVENTS = STATE / "events.jsonl") and consumed by hermes_domain_health.py, hermes_attention.py, hermes_memory_init.py and this weekly distill. Read 2026-10-04: 12.9 MB, mtime 2026-10-03 21:52, 2730 entries in the prior seven days from hermes_domain_health.py (2520) and hermes_cognitive_cycle.py (210). Superseded 2026-10-04: the path this item used to name, ~/.hermes/state/kg_events.jsonl, does not exist, and the writer it named, the ~/.hermes/scripts/kg_ledger/ package, is absent (from kg_ledger.writer import append raises ModuleNotFoundError). fleet_bridge.py and fleet_layer_guard.sh still reference the dead pair, so every kg event write is a silent no-op and the guard reports a missing ledger by construction. See lessons/2026-10-04-the-documented-kg-event-ledger-path-does-not-exist-the-live-.md.
  2. Canonical knowledge graph, ~/clawd/state/knowledge-graph.json (372 nodes/600 edges); engine ~/clawd/scripts/knowledge-graph.py ingests wiki, ~/clawd/memory/YYYY-MM-DD.md, cron jobs+outputs, ~/Projects, HA devices. Weekly rebuild Sun 04:16 (job 8228009396e1); on-demand kg rebuild. Dashboard export ~/clawd/dashboards/graph-data.json.
  3. knowledgelayer, this repo; GitHub-backed, 6h sync, weekly distill.
  4. OpenClaw stores, vector memory ~/.openclaw/memory/main.sqlite (bge-m3 embeddings since 2026-04-15 per backup name), knowledge base ~/.openclaw/data/knowledge-base/ (kb.db + concepts/domains/entities/synthesis), workspace KG event jsonl, memwal extension.
  5. Hermes memories, ~/.hermes/memories/MEMORY.md + USER.md (built-in provider; flush_memories model role configured).

Cross-host shared memory (Pro↔Pro-2): live since 2026-07-09 via the knowledgelayer git leg, Pro is the hub (GitHub ↔ Pro ↔ Pro-2 over the existing SSH channel, 100.97.49.102, pro2 remote with receive.denyCurrentBranch=updateInstead); synced every 6h by knowledgelayer_sync.sh and on demand via kl sync; Pro-2 has its own kl (attribution KL_AGENT=zappo). MemWal/Walrus (engineering-lead-sync) remains an optional richer layer, still blocked on missing MEMWAL_* credentials (names only); plugin installed but disabled. Note: that memo's "GPT-5.5 only / no fallbacks" model policy is superseded, since ~2026-06/07 primary is ollama/kimi-k2.7-code:cloud with explicit fallback chains (see fleet/agents.md).

Log​

  • 2026-07-09: created; superseded 2026-05-01 model-policy note recorded, claude-cowork.
  • 2026-07-09: brain live on Pro-2 (git-over-SSH leg, Pro is hub); authored on Pro-2 as round-trip proof, zappo.
  • 2026-07-09: superseded 2026-05-01 'blocked' status, cross-host sync now live via git leg; MemWal still optional/blocked, claude-cowork.
Sources: ~/.hermes/ops/KNOWLEDGE_GRAPH_STORAGE_MAP.md · ~/.hermes/notes/memwal-context-sync-status.md · ~/.openclaw/data/knowledge-base/