# Memory Scoring v2: Revision-Aware Weighting
## Structural transfer from deception detection to self-improvement
## April 12 2026

### Origin
FM9 revision-probability spectrum was built for deception detection.
Jon challenged: where else does this structure apply?
Answer: my own memory management.

### Problem with v1 scoring
Single formula: 0.45 retrieval + 0.25 recency + 0.20 uses - 0.10 age
Treats all memories identically. This is the warehouse model.

### v2: Revision-class-aware scoring

#### NEAR-ZERO revision class (skills, syntax, historical facts)
Weight: 0.50 retrieval + 0.10 recency + 0.30 uses - 0.10 age
Rationale: stable knowledge, retrieval success matters most

#### LOW revision class (conversation records, established frameworks)
Weight: 0.40 retrieval + 0.20 recency + 0.25 uses - 0.15 age
Rationale: mostly stable, slight recency for recontextualization

#### MEDIUM revision class (framework versions, evolving analysis)
Weight: 0.30 retrieval + 0.35 recency + 0.20 uses - 0.15 age
Rationale: evolution expected, need both old and current

#### HIGH revision class (self-model, task state, preferences, relationships)
Weight: 0.20 retrieval + 0.50 recency + 0.15 uses - 0.15 age
Rationale: current state matters most, old versions are audit trail

### Structural transfer validation
Pattern: classify by revision likelihood, calibrate method to class
Source: evaluating external claims for deception
Target: scoring own memories for retrieval
Shared structure: belief-weighting calibrated to expected change rate