Files
agent-estate-wiki/incidents/2026-11-14-mnemosyne-capture-stall.md
T

16 KiB
Raw Blame History

title, type, status, created, updated, verified_on, review_after, confidence, tags, sources
title type status created updated verified_on review_after confidence tags sources
Mnemosyne capture stalled for 38+ hours (2026-11-14 09:33 AWST onward) — three unclosed alerts, freeze total on re-read incident active 2026-11-15 2026-11-16 2026-11-16 2026-11-18 high
mnemosyne
memory
capture
cron
health-monitor
desktop
pulse
incident
signature-fragmentation
freshness-gate
live-mnemosyne-db
live-cron-executions-db
live-cron-incidents
live-hermes-agent-log
live-mnemosyne-health-monitor-log
live-config-yaml
live-state-db
live-mnemosyne-freshness-gate-script

Mnemosyne capture stalled for 38+ hours (from 2026-11-14 09:33 AWST)

Open. Nothing has been written to Mnemosyne's capture tables since 2026-11-14 ~09:33 AWST, and on 2026-11-15 00:03:35 the Mnemosyne Health Monitor (ea79a2107e36) failed on it (exit 1, incident ea79a2_601037516abd, failure_type: auth, state alerted, unclosed).

Found by the 2026-11-14 nightly audit; not caused by anything in that audit's window (it started the previous morning and simply had not been noticed, because the monitor's own earlier reading of the same condition was classified informational — see below).

Re-read 2026-11-16 00:18 AWST — still open, stall 38 h 46 m, and now three unclosed alerts

Six hours later, at the next nightly audit, nothing has moved. The freeze is total, the monitor has fired twice more, and the scheduler has minted further unclosed rows for the same fault.

  • The counts are frozen, not merely low. working_memory 1130 / episodic_memory 1249 — the same two numbers as every earlier read of the stall (2026-11-15 ~00:4x and 12:03) — so this is not a quiet stretch or a degraded trickle. Zero writes in 38 h 46 m.

  • The whole write path stopped, not just turn capture. Newest rows remain the three [MODEL_REFRESH_PROPOSAL] maintenance writes at 09:32:58 / 09:32:59 AWST; the newest user-turn capture is 2026-11-14 09:04:19 AWST. Maintenance outlived the last turn by ~28 min, then both stopped inside the same minute.

  • cron_incidents 1 → 3 unclosed (138 → 141 rows total). All three are ea79a2107e36, all failure_type: auth, all state: alerted, and each has first_seen_at == last_seen_at — a single observation, no last_seen_at accumulation, no escalation:

    Incident id first_seen_at alerted_at
    ea79a2_601037516abd 2026-11-15T00:03:35.815373+08:00 2026-11-15T00:03:44.559501+08:00
    ea79a2_5f817708701e 2026-11-15T12:03:14.928500+08:00 2026-11-15T12:03:20.245081+08:00
    ea79a2_75303493d484 2026-11-16T00:03:15.418001+08:00 2026-11-16T00:03:22.365399+08:00
  • ⚠ New mechanism, evidenced: the incident signature embeds the elapsed-hours float. The three signatures differ only by 4.64157453 / 4.634034074444445 / 4.643481075555555 — a value never equal twice — so the recurrence never matches the prior row. One durable fault therefore mints a fresh unclosed row every slot, and the estate's unclosed count grows ~1 per 12 h while no single row ever escalates. This job alone now has 15 cron_incidents rows (12 older unknown, these 3 auth). Proposed fix — NOT applied (it changes the monitor's incident path, so it is Anthony's call): normalise the value out of the signature — round it, or omit it — so the same condition re-opens the existing row instead of fragmenting.

The monitor's ≈4.64 h figure is NOT the stall — verified against the gate's own source

Every alert in this sequence prints a near-identical elapsed time while the stall grew from 14 h to 38.7 h. That number is not a measurement of the stall. mnemosyne_freshness_gate.py's newest_user_message() selects the newest messages row with role='user' whose session source NOT IN ('cron','subagent') (the NON_CONVERSATIONAL set, line 52) and returns that row's source; the alert prints (now − that message) / 3600 h. So the figure is the age of the newest conversational message — not the age of the newest capture.

The newest conversational messages on this box are Pulse turns recorded under source: desktop (Pulse delivery check pXXXX-1 …, Pulse morning brief …), on a ~6 h cadence. Each alert's ≈4.64 h lands on the then-newest Pulse turn to within about a minute:

Alert Elapsed printed Now ≈ Newest conversational message it measured
2026-11-15 00:03:35 4.64157 h 00:00:54 2026-11-14 19:22:30 (desktop / Pulse)
2026-11-15 12:03:14 4.63403 h 12:01:25 2026-11-15 07:23:25 (desktop / Pulse)
2026-11-16 00:03:15 4.64348 h 00:01:33 2026-11-15 19:22:59 (desktop / Pulse)

Why they all read ≈4.64: the 0 */12 slots happen to fire ~4.6 h after the most recent Pulse turn. The near-constancy is a cadence coincidence and must not be read as "the stall is only ~4.6 h" — the stall is 38 h 46 m. Same trap class as the naive-UTC note below: a number shaped like a duration that is not one.

Positive proof the capture path is dead rather than merely idle: the identical class of message was captured before the stall — the 2026-11-14 00:09:12 run logged last user msg 2026-11-13 11:21 UTC (desktop) capture delay 75.2 min. Four 2026-11-15 Pulse turns (07:00:43, 07:23:25, 13:23:00, 19:22:59) now have no capture row at all.

The consolidator is alive and every cycle is a deliberate no-op

Latest consolidation_log row 4696 at 2026-11-15T16:14:14.934306+00:00 (= 2026-11-16 00:14:14 AWST), items_consolidated: 0, session_id: auto-consolidate-maintenance; cadence ~31 min (15:12:13 → 15:43:08 → 16:14:14 UTC). New evidence: the row's summary_preview reads No-op consolidation cycle — timestamp updated for watchdog — the cycle touches its own timestamp to satisfy an age test, so a zero-item streak can run indefinitely while every age-based gate reads "fresh". The streak is now ~74 consecutive zero-item cycles, opening 2026-11-14T02:04:42+00:00 (10:04:42 AWST). A zero-item-streak threshold — not an age threshold — is what would have caught this at 10:04 on 11-14.

Adjacent fault in the same runtime — the watchdog lost a slot (2026-11-15 11:31)

96eb95ce12f2 (Mnemosyne Watchdog, every 15 min) failed on its 11:31:20.699 slot, finished_at 2026-11-15T11:49:10.534610 — 17 m 49.8 s, exit 1:

runtime deps guard could not run: Command '['bash', '/home/hermes/.hermes/scripts/hermes-runtime-deps-guard.sh']' timed out after 600 seconds

A 600 s timeout inside a 15-minute slot overran two slots: the 11:45 slot was skipped (no row at all) and the next slot, 12:01:25.120952, completed (→ 12:02:02.657231), so it self-recovered with one slot lost. This is a new signature for this job — the August and October occurrences had the guard quoting another job's error text. Across 11-15 → 11-16 it is the only failed row of its 97. Alert delivered (deliveries.db 11:49:06.052341, for_failure: 1, delivered). Adjacency, not proof of common cause — but the guard and the capture path share the same runtime, so a guard that times out at 600 s belongs next to this stall rather than dismissed.

Symptom (live reads, 2026-11-15 ~00:3x–00:4x AWST)

  • Newest working_memory.created_at: 2026-11-14 01:32:59 — stored UTC, i.e. 2026-11-14 09:32:59 AWST (the row's own timestamp field reads 2026-11-14T09:32:59).
  • Newest episodic_memory.created_at: 2026-11-14 01:33:34 (09:33:34 AWST). Both write paths stopped inside the same minute.
  • Last 8 working_memory writes, all on 11-14 (UTC): 01:01:20, 01:02:39, 01:02:40, 01:03:28, 01:04:19, 01:32:58, 01:32:59, 01:32:59 — then nothing for ~15 hours.
  • Rows by day (UTC created_at): 11-09: 11, 11-10: 10, 11-11: 8, 11-12: 37, 11-13: 14, 11-14: 15. So the pipeline was low-volume but alive before the stall, and the stall is not a normal quiet stretch.
  • Totals at audit time: working 1130 / episodic 1249, DB 27 MB.
  • Gap: at the 00:03:35 alert the newest capture was 14 h 30 m 36 s old; at this audit's read (~00:40) ≈15 h 7 m.

The alert (and its mislabelling)

🚨 Mnemosyne Health Monitor — exception(s):

STALE MEMORY PIPELINE: capture missed: user message 4.64157453 ago (desktop)
has no working_memory row (last capture 2026-11-14 01:32:59)
  • cron_incidents row ea79a2_601037516abd: state: alerted, failure_type: auth, first_seen_at = last_seen_at = 2026-11-15T00:03:35.815373+08:00, closed_at: null. Output file: ~/.hermes/cron/output/ea79a2107e36/2026-11-15_00-03-35.md.
  • ⚠ The auth label describes the classifier, not the fault — now evidenced, not just suspected. The failure is a staleness check and the page records no auth error text. This job's twelve earlier cron_incidents rows read agent ×3, timeout ×1, unknown ×8 (ea79a2_649d8ac46e8a, _f17b45759db0, _573d17e10fc1, _6fb70b032305, _54d6b1f0f58a, _747b27556fd0, _c3ab2d784ba3, _a4661554a3c2, _d93c3acdc4f2, _9197df32320b, _111e67a3d83c, _f317fcd843ae), and the estate-wide table is 53 unknown / 47 agent / 15 timeout / 14 script / 5 auth. agent is also impossible here: the job runs no_agent (plain script). Treat the field as opaque routing metadata.
  • ⚠ The alert text prints the raw stored value (01:32:59, UTC) with no timezone, which reads as a ~23 h gap when the true AWST gap was ~14.5 h. Same class of clock confusion as the monitor's own log line below.

What the earlier slots actually saw — correction, re-derived 2026-11-15 01:2x AWST

An earlier revision of this page got this backwards. It claimed the 12:00:52 AWST run had "already seen the same hole and called it benign" (capture delay -589.5 min), giving the estate ~12 h of unalarmed lead time, and blamed a "sign, not magnitude" threshold. Both claims are withdrawn. The -589.5 min line is not the 12:00 run's — it sits inside the 00:03:35 run's own output block.

Verified by line number in ~/.hermes/logs/mnemosyne-health-monitor.log (1103 lines):

line content
884 === 2026-11-14 00:09:12 AWST health monitor (no_agent) ===
900 ℹ️ last user msg 2026-11-13 11:21 UTC (desktop) capture delay 75.2 min
957 === 2026-11-14 12:01:52 AWST health monitor (no_agent) ===
973 ℹ️ last user msg 2026-11-14 01:03 UTC (ios) capture delay 29.4 min
1030 === 2026-11-15 00:03:35 AWST health monitor (no_agent) ===
1046 ℹ️ last user msg 2026-11-14 11:22 UTC (desktop) capture delay -589.5 min
1047 ❌ STALE MEMORY PIPELINE: …

Two further corrections to the earlier revision:

  • There was no 18:03 run. The job's schedule is 0 */12 * * *; the only 2026-11-14 slots are 00:09:12 and 12:01:52 (both exit 0, delivery suppressed). The "exited 0 at 12:00 and 18:03" sentence is withdrawn — 18:03 belongs to some other job.
  • The detection chain was not broken. mnemosyne_freshness_gate.py fires on magnitude (delay_min < -CAPTURE_LAG_MIN, MNEMOSYNE_CAPTURE_LAG_MIN default 120 min) — there is no sign-only test to blame. The 12:00:52 measurement of +29.4 min is the delay of the newest message that had been captured (the 09:03 AWST ios turn, captured 09:32:59). The gate computes against the newest uncaptured message and only reports a problem when one exists; with the conversation idle from 09:03 to 19:22, every user message still had a capture row, so there was nothing to detect. The first message to postdate the stall was the 19:22 AWST desktop turn (11:22 UTC, the same message the 00:03 alert names); the next scheduled slot after it was 00:03:35, which escalated (-589.5 min = 09:32:59 − 19:22).

Corrected statement of the gap: the pipeline died at 09:32:59 AWST; it became detectable at 19:22 AWST (first uncaptured message) and was detected at 00:03:35 by the next slot — a detection latency of 4 h 41 m after detectability, driven by cadence, not by a broken threshold. What was genuinely unread sits in the consolidation counters: the cycles ran every ~31 min and returned 0 items for 29 consecutive runs (streak opening 2026-11-14T02:04:42+00:00 = 10:04:42 AWST; last productive cycles 4622/4621/4620 at 09:33:34 / 09:33:16 / 09:01:40). Every run printed Last consolidation: … (0 items) — including the "healthy" 12:00:52 one — but the gate tests consolidation age only (> 36 h, default), never the item count. An empty-but-current cycle is the loudest signal available and the only one nothing watches.

Impact

Magnitude superseded 2026-11-16 00:18 AWST: 38 h 46 m uncaptured (was ~15 h), with two more unclosed alerts and four 2026-11-15 Pulse turns (07:00:43, 07:23:25, 13:23:00, 19:22:59) added to the uncaptured set. The ~15 h figures below stand as the 2026-11-15 readings.

  • ~15 hours of context is not in memory: the desktop/Bot Chat session that ran to 19:22 AWST (the last user message the monitor could see), all Telegram traffic, and every cron session in the window. Recall over that period will answer from nothing.
  • The 23:30 AWST sleep/consolidation cycle ran and found 0 items to consolidate (Last consolidation: 2026-11-14T15:30:51 UTC (0 items)), so consolidation is not compensating.
  • Nothing else in the estate broke: 2026-11-14 settled 167 rows — 166 completed / 0 failed / 1 running in executions.db, deliveries 4 / 4 delivered, and no other cron row went red. This is a memory-path incident only. The one exception is the monitor's own job: it is now the fleet's single non-clean row (last_status: error, failure_streak 1, previously ok / 0) and cron_incidents went 0 → 1 unclosed (138 rows total). 2026-11-15 so far: 8 rows — 7 completed / 1 failed (this job), 1 delivery delivered.

Not the known config killer

The previously recorded capture-killer on this host is a stringified list/dict written into ~/.hermes/config.yaml by hermes config set. Checked and clean:

  • config.yaml mtime 2026-11-13 18:03 — the file was not modified during or before the stall began.
  • A full YAML walk found no scalar that begins with [ or {; memory.mnemosyne.* parses as proper lists (ignore_patterns: [], skip_contexts, sync_roles).

Open / unverified

  • Root cause: not established. No capture-path error lines appear in agent.log after 09:33 — the writes simply stop. Candidates to test next: a stuck capture worker or watermark, the pulse/heartbeat traffic that ran 12:03–19:22, session-side memory_enabled / mnemosyne.profile_isolation interaction, or the sleep cycle's consolidation_claimed_at state.
  • ✅ Resolved — the 00:03 escalation DID deliver. deliveries.db holds exactly one 2026-11-15 row: 2026-11-15T00:03:39.052888+08:00, status: delivered, for_failure: 1 — i.e. 3 s after the incident row was raised and 5 s before its alerted_at (00:03:44.5). The notification path worked; the detection path did not (see the ℹ️-at-12:00 bullet on systems/scheduled-tasks).
  • ✅ Answered 2026-11-16 00:18 — capture did NOT resume. Two further monitor slots failed on the same condition (12:03:14 and 2026-11-16 00:03:15) and the row counts are unchanged (working 1130 / episodic 1249). Closed as no; the underlying cause remains open.
  • The estate's only automated alarm is understating this incident. The alert's ≈4.64 h is a cadence artefact, not a stall duration — anyone reading the alert text alone would conclude the pipeline is ~4.6 h behind when it is 38 h 46 m. A stall-duration figure (newest capture vs now) would fix it.