Files
agent-estate-wiki/systems/mnemosyne.md
T

14 KiB
Raw Blame History

title, type, status, created, updated, verified_on, confidence, tags, sources
title type status created updated verified_on confidence tags sources
Mnemosyne system active 2026-07-22 2026-10-18 2026-10-18 high
memory
mnemosyne
backend
live-probe
mnemosyne-db
config
ops-profile-config

Mnemosyne

Purpose

Primary agent memory store for Hermes.

Current Configuration

  • Provider: mnemosyne
  • Enabled: true
  • Auto sleep: true
  • Default scope: session
  • Memory char limit: 5000
  • User char limit: 3000
  • Vector type: int8
  • Profile isolation: true
  • Shared surface read: true
  • Shared surface path: data/shared/mnemosyne.db
  • Skip contexts: cron, flush, subagent, background, skill_loop
  • Reflect max calls per session: 3
  • Sync roles: user
  • Flush every: 6 turns
  • Embedding model: BAAI/bge-small-en-v1.5 (384 dims)
  • Embeddings via API: false (local fastembed)

Database Stats (live, 2026-10-04)

Table Rows Notes
working_memory 773 Active memories
episodic_memory 1,047 Consolidated summaries
memory_embeddings 760 Working-tier embeddings
memory (legacy) 2,181 Old table, not migrated
vec_working / vec_working_rowids 760 Working vector index
vec_episodes / vec_episodes_rowids 1,035 Episodic vector index
vec_facts_rowids 0 memoria_facts is intentionally left unvectorised
Shared surface 0 Empty — cross-agent sharing not functional

Embedding coverage (live, 2026-10-04): 1,759/2,194 memories embedded (80.2%) — working 726/752 (96.5%), episodic 1,033/1,039 (99.4%); 13 embeddings generated in the previous 24h. Working coverage is above the 95% alert threshold enforced by the Embedding Coverage job (b63ed5817212, every 6h, Telegram topic 102).

Known Issues

Tier counts (live, 2026-10-09 22:27): working 773, episodic 1,047, canonical_facts 97, facts 871, memoria_facts 407, memoria_preferences 132, conflicts 152 (1 open — id 152, contradiction, created 2026-10-09 13:11:38), working embeddings 760/773, episodic embeddings 1,043/1,047 (99.6%). Ranking relevance is the live issue, not coverage — see incidents/2026-10-09-memory-recall-relevance.

Tier counts (live, 2026-10-12 12:02 — health job ea79a2107e36, green): working 760/786 embedded (96.7%), episodic 1050/1056 (99.4%) — both above the 95% floor, no reindex needed (b63ed5817212 continues to backfill). Orphan embeddings 0. FTS5 exactly aligned: fts_working 786/786, fts_episodes 1056/1056, fts_facts 876/876. 0 unconsolidated working rows; 1 unresolved conflict (threshold 10); last consolidation 0.3 h earlier; 2,252 memories total (786 working / 1,056 episodic / 1 triple); disk 15 GB free (81% used). Interpretation: coverage is not the live issue — ranking relevance is (see incidents/2026-10-09-memory-recall-relevance).

Tier counts (live, 2026-10-12 00:1x): working 786, episodic 1,056, facts 876 (fts_facts 876 — FTS aligned), fts_working 786, memoria_facts 410, conflicts 152. Embeddings: working 773/786 (98.3%), episodic 1,052/1,056 (99.6%) — both above the 95% reindex threshold; the coverage job b63ed5817212 backfilled 15 vectors since the 2026-10-11 00:08 check.

Capture clock is naive UTC — never compare it to AWST now (2026-10-11 false alarm). At 2026-10-11 00:08 the health monitor raised an EXCEPTION claiming the capture path had stalled for 10.7 hours, citing working_memory MAX(created_at) = 2026-10-10 13:24:53 against the local clock. created_at is stored as a naive UTC string (the same trap as consolidation_log), so the real gap was 2.7 h: that write landed at 21:24:53 AWST, inside the 21:13 session that produced it. Capture was healthy throughout — working_memory gained 2 rows on 2026-10-11 (11:19:37 and 11:51:39 UTC = 19:19 / 19:51 AWST), both from the day's only interactive session. Convert to local before judging staleness; a raw MAX-vs-now comparison always looks ~8 h stale in AWST.

Vector embedding coverage gap — RESOLVED (2026-08-02)

  • Was: only 159/1,476 working memories (10.8%) embedded and 0 episodic, so recall fell back to FTS5 for ~90% of working and all episodic memories.
  • Resolved by the 2026-08-02 reindex pass. Live coverage on 2026-10-04 is 96.5% working / 99.4% episodic, and new embeddings are still being generated. The figures above are kept as the before-state only.
  • Standard repair if coverage drops below 95% working: ~/.hermes/hermes-agent/venv/bin/mnemosyne reindex --yes (the Embedding Coverage job b63ed5817212 triggers the same path every 6h).

Vector recall silently ran FTS-only (2026-10-03 — repaired)

  • fastembed was absent from every uv-managed runtime, so the dense leg of mnemosyne_recall produced no candidates and recall fell back to FTS5/keyword scoring without raising an error. Sessions kept working, which is why it went unnoticed.
  • Compounded by FASTEMBED_CACHE_PATH being unset in ~/.hermes/.env, and by the embedding worker sometimes launching in isolated (python -I) mode, which does not inherit PYTHONPATH.
  • Repaired 2026-10-03: fastembed installed into the bootstrap-selected runtime, FASTEMBED_CACHE_PATH pinned, and hermes-runtime-deps-guard.sh extended to assert fastembed + cache path + model dir every 15 min alongside its existing dotenv/mnemosyne checks.
  • CLI caveat — do not misread mnemosyne doctor. Doctor and the recall CLI run under the legacy venv/ (Python 3.11), where the sqlite_vec extension is not importable; doctor therefore prints vec0: present_but_unloadable and marks the vector tiers unavailable even while the gateway embeds normally. Judge vector health with ~/.hermes/scripts/check_embedding_coverage.py, not with doctor's vec0 line.
  • Benign noise: pthread_setaffinity_np failed … mask: {0,} from onnxruntime in mcp-stderr.log is a container affinity warning emitted on every model load — not an error.
  • See incidents/2026-10-03-mnemosyne-vector-recall-fts-only.

Profile isolation chosen over shared surface

|- The ops profile was enabled with mnemosyne on 2026-08-02 |- Decision: profile_isolation: true was set on ops, giving each profile its own independent memory bank — zero cross-contamination |- data/shared/mnemosyne.db exists with 0 rows; cross-profile sharing is intentionally unused |- This means Tusk (ops) and Rhino (default) do not share memories, which is the desired behaviour |- Config in ~/.hermes/profiles/ops/config.yaml: memory.mnemosyne.profile_isolation: true, allow_tool_override: false, shared_surface_read: false |- Verified end-to-end on 2026-08-02: memory writes land correctly in both profiles independently

Legacy table not migrated

  • 2,181 rows in the old memory table are not in the new working_memory/episodic_memory system
  • These are only accessible via the legacy memory tool, not via mnemosyne_recall

Gateway notifications were captured as user memories (fixed 2026-09-30)

  • Hermes's background-process notifications (long-running terminal jobs reporting completion) were being written into Mnemosyne as [USER] [Anthony Martin] memories — i.e. machine messages were stored as if the user had said them, and then surfaced during recall as if they were his words.
  • 36 such rows were found and purged on 2026-09-30 (working tier plus the corresponding episodic rows). Recall after the purge returned clean.
  • Two canonical facts were also wrong because of the same class of drift: hermes_cron_model_default named the removed provider custom:omni instead of custom:omni-101, and the Daily Marketing Job Radar's model fact was stale. Both corrected.
  • Rule: treat a recalled memory attributed to a person as untrusted until the attribution makes sense in context. Machine-generated status text should never enter the person-attributed tier.

Known Management Jobs

  • Mnemosyne Auto-Consolidate every 30m — script: consolidate_mnemosyne.sh, mode: no-agent
  • Mnemosyne DB Backup weekly Sunday 03:00 — script: mnemosyne-backup.sh, mode: no-agent
  • Mnemosyne Health Monitor every 12h — skill: mnemosyne-operations, reports only on exception
  • Mnemosyne Embedding Coverage every 6h — job b63ed5817212, alerts to Telegram topic 102 only when working-tier coverage drops below 95%
  • Mnemosyne Watchdog every 15m — script: mnemosyne-watchdog.py, mode: no-agent
  • ⚠ Blast-radius note (2026-10-16). Both monitors above (ea79a2107e36 Health every 12 h, b63ed5817212 Embedding Coverage every 6 h) are pinned to custom:omni-101 — the same node the memory system's own reporting depends on. During the 2026-10-14 → 10-15 omniroute-1 outage both failed on every slot, so no coverage figures were refreshed for ~34 hours; the last explicit read remains the 2026-10-12 12:02 run (working 760/786 = 96.7 %, episodic 1050/1056 = 99.4 %), and the 2026-10-16 00:09:48 Health Monitor run is the first clean read since. The two script-only jobs (consolidate_mnemosyne.sh, mnemosyne-backup.sh) were unaffected — they make no model call. Monitoring that shares a blast radius with its subject is degraded monitoring; a coverage alert that cannot fire during a router outage is exactly the case this has to cover. → incidents/2026-10-14-omniroute-node-connection-refused-storm

History and Evidence

  • 2026-10-03: vector recall found running FTS-only (fastembed missing from the managed runtimes); repaired the same day and the dependency guard extended — see above and incidents/2026-10-03-mnemosyne-vector-recall-fts-only.
  • 2026-09-30 memory-hygiene pass: 36 gateway-notification rows purged from the person-attributed tier; two canonical facts corrected (hermes_cron_model_default → custom:omni-101, job-radar model). No unresolved conflicts left in the working tier.
  • 2026-08-19 Mnemosyne Watchdog incident: Watchdog job (96eb95ce12f2) cycled for ~2 hours early morning Aug 19 reporting no consolidation for 36–41.5+ hours (last consolidation Aug 17 07:15). Watchdog stopped after ~2 hours of alerts. See systems/scheduled-tasks Known Issues section.
  • 2026-08-18 incident: Working memory capture (sync_turn) stopped functioning after 2026-08-16 23:15:39 — ~48 hours of conversation turns not captured. Root cause: silent guard conditions in run_agent.py:_sync_external_memory_for_turn(). Config is clean (native YAML types). MEMDEBUG logging recommended to identify which guard blocks. See incidents/2026-08-18-mnemosyne-capture-pipeline.
  • 2026-08-01 incident: Automatic conversational capture was broken by stringified structured YAML, especially sync_roles: "['user']". Fixed by rewriting with native YAML types. See incidents/2026-08-01-mnemosyne-capture-failure.
  • 2026-08-01 cleanup: Legacy always-injected memory reduced from 8,013 bytes to 4,998 bytes (memory) and 2,958 bytes (user).
  • 2026-07-29 incident: memory.mnemosyne.tools: None caused schema loading warnings. Fixed by removing the line. See incidents/2026-07-29-mnemosyne-relay.
  • 2026-07-29 relay scope corruption: RuntimeError: scope handle is not at the top of the stack — 3 occurrences, self-resolved.

Re-verified 2026-10-10 (audit window; live probes 2026-10-11 00:14-00:20 AWST)

  • Gateway load path repaired. The gateway's activate_dependencies() selects one committed uv env (795020dbf24343cd9c52df24723dca83); it was missing mnemosyne / mnemosyne_hermes and the native dependencies. Repaired 2026-10-10 21:26-21:34 and confirmed live from 22:01:19 (registered (40 tools) -> activated). Detail: incidents/2026-09-28-hermes-managed-runtime-deps.
  • "Capture silent since 10-09 15:20" was a false alarm. It resolved to no interactive sessions in the window - the window was cron-only traffic. What the surrounding hours did capture was dominated by gateway background-process notifications stored as [USER] [Anthony Martin] [IMPORTANT: Background process ...] rows: recall-visible, low signal. A capture filter would help; no fix applied (carried open item).
  • Coverage script executable again: ~/.hermes/scripts/check_embedding_coverage.py now carries mode 0755, so the monitor's non-executable finding is resolved.
  • Open, unverified: 13 orphan rows in memory_embeddings (oldest 2026-08-22) reported by a 2026-10-10 monitor run - neither confirmed nor cleaned by this audit.

Re-verified 2026-10-18 (audit window; live probes 2026-10-18 00:30-00:36 AWST)

  • Both monitors rode the omniroute-1 503 window; both are green again. ea79a2107e36 (health, every 12 h) failed 10-17 00:00 with RuntimeError: HTTP 503 ... resource pressure and delivered its alert; b63ed5817212 (coverage, every 6 h) failed 10-16 12:03 the same way. Since 10-17 12:00 every slot of both has completed with delivery_outcome: suppressed (nothing to report) — including 10-18 00:00.
  • No fresh coverage figures exist in the ledger. These jobs report to Telegram, not to the execution row, so the last explicit read remains 2026-10-12 12:02 — working 760/786 (96.7 %), episodic 1050/1056 (99.4 %), both above the 95 % reindex threshold. Nothing here indicates a reindex is due.
  • The watchdog triple-checks the same thing. 96eb95ce12f2 (every 15 min) exited 1 four times — 10-16 12:15 + 12:30, 10-17 00:15 + 00:30 — each row quoting the coverage monitor's own 503 error, i.e. one sentinel producing four failure rows. Read failures per job, not per row. It has been clean on every slot since (latest 10-18 00:30).