dotfiles/claude/memory/project-mempalace-upgrade-3-4-0-plan.md

name, description, metadata

name	description	metadata
project-mempalace-upgrade-3-4-0-plan	Scheduled MemPalace 3.3.6 → 3.4.0 upgrade — approach ratified 2026-06-07, execute a few days later once past the heavy ARC work. Backup first; migrate-wings and repair explicitly excluded.	node_type type originSessionId memory project 5a8200e7-d734-4dc1-b69e-93d77076cab2

MemPalace upgrade 3.3.6 → 3.4.0 — scheduled plan (ratified 2026-06-07)

Steward directive: execute "in a few days once we've gotten past the heavy work" (the ARC SCSS day 2026-06-08 + whatever follows). Never change the memory substrate the night before — or morning of — a big work day.

Why upgrade: 3.4.0 (92 commits, on PyPI) ships a filtered-search fallback (#1245)

• MCP stale-index retry (#1635) — converts our most-hit failure (#1665 wing-filter Error finding id) from hard error to graceful degradation — plus diary_write/hook repairs (#1693, #1700: content alias, −32602 diagnostic) that touch the exact machinery our Stop/PreCompact hooks fire.

Why caution: our two filed issues remain OPEN — #1665's root cause (HNSW drifts from sqlite after bulk delete) is unfixed, and #1589 (repair rebuilds an index chromadb SIGSEGVs on) is untouched. Palliatives, not cures. Full read: CM-AI docs/thinking/David/l1-reliability/substrate-failure-lessons-2026-06-07.md §10.

The sequence (one sitting, ~30 min)

1. Backup the palace first — ~/.mempalace/palace-memory (and the typography palace if touched): same-disk copy, verify count. No backup → no upgrade.
2. Upgrade the installed tool to 3.4.0 (installed via pipx/uv — check which mempalace for the carrier; currently 3.3.6).
3. Re-apply / verify the local recency-ordering fix in layers.py — the wake-up CLI's handoffs ordering depends on it; if the upgrade clobbers it, the thread-lineage glance reads oldest-first (wake §2.b.0.5 documents the symptom).
4. Re-test wing-scoped search (mempalace_search with wing:) — expect graceful fallback now, not Error finding id. Either way keep the unscoped+post-filter workaround until #1665's root cause closes.
5. Smoke the session protocol: mempalace_status, one kg_query, one diary_read, CLI wake-up --wing claude-sessions (ordering check per step 3).
6. Do NOT run migrate-wings (wing-name normalization — interesting for our odd wing names but a 61k-drawer migration wants its own considered session, after a verified backup and a read of the migration guide). Do NOT run repair (#1589, still booby-trapped).
7. Note the outcome in the day's ledger + this file; if the wing filter now degrades gracefully, update feedback-mempalace-wing-filter-broken.

Same slot — our own memory system's maintenance (added 2026-06-07 night)

Three bounded improvements to the files-authoritative layer, ratified in principle, to execute in this same post-heavy-work sitting:

1. MEMORY.md archive pass — the index is over capacity (531 lines / 267 KB; the harness loads it PARTIALLY every session with a warning). Our own silent-truncation exposure: move Archived-section entries older than ~2 weeks into MEMORY-archive-2026.md (linked from the index), keep the index to current threads + standing preferences. Top-of-file loads first, so the fresh entries are safe meanwhile.
2. Wake canary (skill-harvest proposal, steward to authorize) — apply lesson 7 to ourselves: a cheap wake/wrap probe — every [[link]] and index pointer in MEMORY.md resolves to a file; flag dead pointers and (optionally) load-bearing orphans. One script, runs in seconds, catches our index drifting from our store.
3. Palace backup as standing practice — not just before the upgrade: a periodic same-disk copy of ~/.mempalace/palace-memory (launchd or wrap-step), because "restore is the only test of backup" and the palace currently has none.

Standing posture regardless of version: witness, not notary — per the 2026-06-07 CLAUDE.md amendment (conflicts = verification triggers). MemPalace remains the unaffiliated stopgap; BMF is the destined replacement (mempalace-is-unaffiliated-stopgap).