rdmsm4x changelog — overnight fleet health: mem0 truth, mem0 caching, Tyrell Build 33

One line: Tyrell's "Mem0 cannot connect" on spokes was a dashboard defect, not an outage. It is fixed and shipped in Build 33 to 5/6 hosts, spoke silo memory daemons are retired, and hub memory search now caches (repeat 85 ms → 0.26 ms, model resident 24 h).

• When: 2026-09-26 00:02:39 → 2026-09-26 00:56:27 EDT
• Host / agent: rdmsm4x · claude@rdmsm4x/3f92c388 (Opus 5.5 lead; 2 Sonnet workers for the repo and service sweeps)
• Ticket: EPIC-20260926-01 (children TASK-20260926-02/-03/-04); resolved TASK-20260924-23, ISSUE-20260924-30, TASK-20260924-14

Scope (hosts)

rdmsm4x, rdmbair15m5, jdmbair13m5, rdmpw3265m, rdmpw3275m. rdmbair13m5 was offline all night (Tailscale last seen ~4 h at 00:12).

Changes

1. mem0 diagnosis: agent memory worked from every reachable spoke (MCP mem0_health through ~/bin/mem0-mcp → hub, 27 vectors). The Tyrell dashboard read a Qdrant path that doesn't exist (Build 32), or would have read an empty local silo ecsmem0d (main 0e7591e).
2. Tyrell main 24321a8, a144a1d, fa36509, 9817f7a, 250b8db, 1313eee (Build 33 stamp 1e37a55, tag v0.2.0-build33). Hub tyrelld serves read-only /api/fleet/mem0/{health,v1/telemetry}. Spoke dashboards read it. Spokes run no ecsmem0d, and existing plists are moved aside at onboarding and on every launch. Also fixed two pre-existing test breaks.
3. Tyrell Build 33 notarized 82044f4c, stapled, DR unchanged, universal2. Canary rdmbair15m5, then fleet, then hub. Measured app 33 + running daemon-45bc0b545e92 on 5/5 reachable hosts.
4. Spoke silos retired 00:21 on rdmpw3265m, rdmpw3275m, rdmbair15m5, jdmbair13m5. launchctl bootout, plist moved to ~/Library/Application Support/Tyrell/retired-launchagents/, data untouched.
5. ECSMemory main b067fd9 + bc47aff: exact-text embedding LRU (ECSMEM_EMBED_CACHE_SIZE=512), keep_alive 24h, and cache telemetry. 329/329 tests. Deployed to the hub as pinned releases under ~/Library/Application Support/ECSMemory/releases/b067fd9cbbeb and Tyrell/releases/ecsmem0d-b067fd9cbbeb; ecsmem0d kickstarted.
6. Repos: sweep of 206 repos: 162 clean/synced, 5 fast-forward pushes, upstream tracking set on 36 (fleet/overnight/repo-upstream-set-20260926.log), 2 stale 40 h index.lock moved to ~/dev/_quarantine/stale-index-locks-20260926/, and uncommitted docs/todo work snapshotted (~/.agent-coordination/snapshots/rdmsm4x/{docs,todo}/20260926-002006).
7. Old Tyrell copies: 2 empty root-owned rollback dirs moved to ~/archive-quarantine-20260926/app-revisions/ (rdmsm4x, rdmbair15m5) with MANIFEST.tsv.

Verification

• Tyrell swift test on rdmsm4x: 1019 Swift Testing / 0 failures (Build 32 gate 984); XCTest 423 (+184, +11), 0 failures. The negative control fails as expected.
• ECSMemory: 212 unit + 117 E2E = 329, 0 failures.
• Relay from rdmpw3265m returns vector_count 27, commit b067fd9, relay tyrelld; /v1/memories → 404 (not relayed).
• Canary 12-min watch: same pid, 30–79 MiB, 24/24 port probes 200.
• Not verified: a visual screenshot of the spoke dashboard.

Undo

• Tyrell: /Applications/.tyrell-rollbacks/<20260926-…>/Tyrell.app (Build 32) per host via install_app_local.zsh, plus the previous daemon release under ~/Library/Application Support/Tyrell/releases/daemon-098e1b457bc4.
• ECSMemory hub: targets in ~/dev/fleet/overnight/ecsmemory-deploy-rollback-20260926.txt (ln -sfn), then kickstart com.eastcoastscience.ecsmem0d.
• Spoke silo: move the plist back from retired-launchagents and launchctl bootstrap gui/$(id -u) . Build 33 will retire it again.

Outstanding

• rdmbair13m5: needs someone to check the machine physically. Build 33 install is tracked in TASK-20260926-04.
• Ollama down on both Intel hosts: TASK-20260926-02. GitHub backup gaps (ai/LLM archived, llm-wiki GH007): TASK-20260926-03.
• FEAT-20260924-13 (ecsmem0d over the tailnet) still needs Rich's acceptance of the tailnet exposure. Nothing changed there tonight.