session 2026-07-07: MemPalace wind-down decided (evidenced audit) + KG exported to files

Diagnosed the "upgrade" as a 300-behind bge-m3 fork, not a standard install; the
morning "self-healed" drift was 108 chronic quarantines (data intact). Independent
46-transcript audit: search+KG not load-bearing (value migrated to files). Steward
decided (evidenced) to wind down palace-memory; execute fresh. KG exported to
knowledge-graph.jsonl (329 triples + 437 entities). Wind-down plan + BMF-lessons
report outline captured. Demote-on-promote; skill harvest appended.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dsqe1x23NgWaRdiFWSotur

claude/memory/project-mempalace-winddown.md

@@ -0,0 +1,45 @@
+---
+name: project-mempalace-winddown
+description: "Decision (steward, evidenced, 2026-07-07): wind down the palace-memory MemPalace instance and rewire wake/wrap to the files layer. Execution deferred to a fresh session. KG already exported + secured. Typography palace KEPT, untouched."
+metadata: 
+  node_type: memory
+  type: project
+  originSessionId: 9f112d92-72f5-434e-8b7f-cde0e80da021
+---
+
+# MemPalace wind-down — palace-memory
+
+**Decision (steward, 2026-07-07):** wind down the `palace-memory` MemPalace MCP instance. It is the unaffiliated stopgap ([[mempalace-is-unaffiliated-stopgap]]); BMF/CapableMind is the destined replacement. Execute the rewrites **fresh/rested**, not at the tail of the long 2026-07-07 session.
+
+## Why (evidenced — full detail in `session-ledger-2026-07-07.md`)
+Independent audit of 46 sessions: MemPalace's **search + KG are NOT load-bearing** for palace-memory — confirmatory/redundant. Three convergent tests: (a) `find_tunnels`/`traverse` used **0× ever**; 27/31 `kg_query` = a byte-identical wake ritual; genuine reach ~0.43/session, declining to ~0.07 in July; (b) 5/5 distinctive search payloads found verbatim in the Markdown; **zero** transcript statements credit search/KG with file-absent content; (c) write:read ≈ 7:1; 437 entities all type 'unknown'. Heaviest use (06-30 clasp) returned a **stale claim the files corrected**. Confidence: counts 0.93, verdict 0.82. Cost of keeping: HNSW↔sqlite drift (#1665/#1589) fired live at the 07-07 wake — silent-bad-recall class, **no safe repair path**; install is an editable fork ahead 5/behind ~300.
+
+## DONE (2026-07-07 — no-regret, reversible)
+- **KG exported + secured:** `memory/knowledge-graph.jsonl` (329 triples) + `memory/knowledge-graph-entities.jsonl` (437), all JSON-validated. Git-tracked (dotfiles, dual-remote) → strictly more durable than the SQLite KG. Highest-value slice = 84 `drift-pattern` triples (most already mirrored in feedback-memory files).
+- **Verified backup:** `~/.mempalace/palace-memory.pre-rebuild-bak-20260707-1230` (chroma 88,159 + KG). Keep until wind-down verified complete.
+- Throwaway partial re-embed dir removed.
+
+## STAGED PLAN (execute fresh)
+1. **Stop feeding it** — disable the Stop/PreCompact hooks that fire `diary_write`/`add_drawer`/`kg_add` (cheap, reversible; `mempalace hooks` / settings).
+2. **Drop the dependency** — remove the `mempalace` MCP server from the Claude Code config so it stops loading + quarantining drift every wake.
+3. **Rewire skills to files** — `/wake-up` and `/wrap-up` currently query MemPalace heavily; point them at the Markdown layer (`MEMORY.md`, `session-*.md`, `REVIEWED/PENDING`) + `knowledge-graph.jsonl` instead. Wake reads the drift-patterns from the jsonl (replaces the `kg_query claude-code` ritual). Bounded skill edits; run the skill-harvest register for the proposals.
+4. **Cold-archive** the backed-up palace; delete the live churning one after verification.
+
+## PRESERVE — do NOT touch
+- **Typography palace** (`~/.mempalace/palace-chamber-typography`, queried via CLI for ARC type work, [[reference-typography-palace-cli]]) — SEPARATE instance, genuinely used, KEEP as-is. The audit verdict is palace-memory ONLY. Do **not** do a global embedder flip (would break it, 1024-dim/bge-m3).
+- The exported drift-patterns — wire wake to read them.
+
+## BMF lessons report — OWED (steward-requested 2026-07-07), part of decommission (tomorrow)
+**Framing (steward's own):** MemPalace was a long experiment — stopgap AND *reconnaissance*: a substrate close to our own vision, run to see **from the inside** how others approach the AI-memory problem. The report harvests everything of value **FOR BMF** (CapableMind L1, the real replacement), not a mere obituary.
+**Home:** `~/_Dev/CapableMind-AI/docs/thinking/David/l1-reliability/mempalace-decommission-report-for-bmf-2026-07-07.md` — build ON `substrate-failure-lessons-2026-06-07.md`, do not duplicate.
+**Sources to gather (fast):** substrate-failure-lessons-2026-06-07.md + sibling l1-reliability post-mortems (recall-quality-baseline, ingestion-forensic, deep-audit-plan); MemPalace's own `~/_Dev/mempalace/CLAUDE.md` design principles; THIS session's forensics (`session-ledger-2026-07-07.md` + the audit verdict); PENDING-16 (observation-recall coupling) + PENDING-17 (epistemic integrity) = BMF design intent to map to; the MemPalace feedback memories.
+**Drafted outline (my synthesis this session — fill this in):**
+1. *What MemPalace is* (the idea close to ours) + *what we ran* (~4 mo, ~88k drawers, session substrate alongside a files layer).
+2. *Verdict — delivered vs aspirational:* search+KG NOT load-bearing (audit: ~0.43 reach/session→~0.07; graph features 0×; write:read 7:1); value **migrated to files**; the "cheap context / 100% recall" promise vs the fragile reality.
+3. *Failure modes → BMF lessons:* HNSW↔sqlite drift, no safe repair → index must be **rebuildable from an authoritative store** + drift **detectable/auto-healed** (validates L1 logchain-as-truth); silent-degradation-quarantine → **honest degradation** (constitutional constraint #4 — MemPalace = case study of the violation); repair-SIGSEGV → **recovery is the most-tested path**; heavy embedder + global-config lock-in + 21h re-embed → **swappable/deterministic-re-embed-from-source, per-collection**; write:read 7:1 bloat → **validates PENDING-16 observation-recall coupling**; fork-maintenance (300 behind) → **own the stack**.
+4. *Ideas worth carrying:* verbatim storage; **temporal-validity KG** (valid_from/to, invalidate); hooks-driven zero-token background ingestion; cross-session **diary/continuity**; local-first/privacy-by-architecture; a scannable compressed index (AAAK) — concept worth evaluating. **Decorative-not-load-bearing** (avoid): the wing/room/drawer loci metaphor (wing-filter broken, entities all type 'unknown', graph-traverse never used).
+5. *THE deep lesson — files-authoritative convergence:* value migrated to git-versioned Markdown; the durable substrate should **BE** the verifiable/versioned layer, with vector/graph a **rebuildable-from-truth index, not the primary store.** Maps to L1 logchain-as-truth + PENDING-17 (don't launder uncertainty).
+6. *Decommission record:* KG exported ✓; typography palace kept; backup retained.
+
+## Moot now
+- The 3.3.6→3.5.0 fork-rebase / embeddinggemma-vs-bge-m3 question is **moot for palace-memory** (we're leaving it). Typography stays on bge-m3 as-is unless separately revisited.