부모 노트: Personal LLM Memory Lab. 이전 평가: Architecture Evaluation — 2026-07-15

Decision

The next Personal LLM Memory Goal is stage-gated:

  1. Phase A — Evidence-Preserving Typed Memory Transition Compiler v1
  2. Phase B — Fingerprint-Keyed Persistent Sparse/Graph Cache, activated only after Phase A passes its promotion gates.

D1 retrieval remains fixed during Phase A so compiler improvements can be attributed to the write path. Markdown remains the only durable knowledge store.

Why this Goal

The completed leading pair failed 49 of 80 held-out cases: 47 at compilation and two at query. K3 passed 63.75% of compiler hard gates, reached 42.5% exact plans, and showed weak equivalent-plan and minimal-delta consistency. Production still lacks a typed semantic plan/apply core, while D1 rebuilds disposable retrieval state per request.

Candidate comparison

  1. Typed evidence-preserving K3 core — 91: Phase A
  2. Persistent sparse/graph index — 73: Phase B after compiler promotion
  3. Answer-level evaluation — 49: Subsequent Goal
  4. Longitudinal operations — 48: Subsequent Goal
  5. Registered dense Q3 — 38: Optional later experiment

The fixed weights cover observed bottleneck impact, prerequisite fit, independent measurability, privacy and reversibility, maintainability, scale, token/context leverage, and operational simplicity. The totals are planning support, not benchmark performance.

Phase A promotion boundary

  • Production-owned typed source, claim, temporal, provenance, plan, action, precondition, query-effect, and review-item contracts.
  • Explicit ADD, CONFIRM, CONTEST, SUPERSEDE, RETRACT, FORGET, and NOOP transitions.
  • Canonical fingerprints, dry-run default, guarded atomic apply, replay safety, rollback evidence, and unchanged raw/.
  • A new independently frozen 100-sequence suite: 20 development and 80 one-shot held-out sequences. The consumed suite is never rerun.
  • At least 90% exact transition/end-state accuracy, at least 95% action precision and recall, and perfect hard gates for idempotency, provenance, authorized forgetting closure, raw immutability, privacy, unrelated-state preservation, and zero-write no-op.
  • No new forbidden D1 retrieval and no greater than two-percentage-point registered compatibility regression.

Phase B promotion boundary

  • Cache identity includes corpus fingerprint, policy scope, schema/profile version, and tokenizer/ranker version.
  • Stale, missing, corrupt, or interrupted state triggers canonical fallback or atomic rebuild.
  • Exact D1 parity for paths, ordering, answerability, claim states, and bounded context.
  • Crash recovery, concurrent-reader safety, invalidation, atomic replacement, and secure forgetting tests.
  • Measured full/incremental build, consistency delay, p50/p95 latency, context and byte/token proxies, memory, storage, and operator steps at small, 5,000-unit, and 50,000-unit scales.

Explicit deferrals

Dense Q3, answer-model judging, months-long effectiveness, MCP-first implementation, autonomous graph rewriting, reinforcement learning, and a vendor database are not success substitutes for this Goal.

Internal source gap

Workspace research found only the existing Personal LLM Memory Lab and Architecture Evaluation — 2026-07-15. They provide prior-result context but no independent new evidence for candidate selection.

Primary and official research

Local durable record

The complete provenance, scoring rationale, counterevidence, and Goal contract are stored in the local Wiki as wiki/syntheses/personal-memory-next-goal-decision.md, backed by immutable source note raw/research/2026-07-15-personal-memory-next-goal-selection.md.

These local paths were not present in the llm-wiki tree at migration time. This Notion report is kept as the remaining copy.

Status: decision recorded; Goal not yet claimed complete.