P327: Fortunate Recall
Mullick, Ansuman; Tuzun, Eray · arXiv:2609.10413 · live #pattern-327 · commit 1d0b2d4
Full title: Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs (9 Sep 2026). Current LLM memory systems treat every personal fact the same — ethnicity, a current food preference, and a six-month-old dentist appointment share identical retention and retrieval priority. Stores grow without bound; precision degrades; confabulation rises.
Fortunate Recall (FR) reframes the problem as lifecycle management, not retrieval. A 10+1 behavioral ontology (Health and Wellbeing, Logistical Context, Preferences, Relationships, and siblings — behavioral domains, not cognitive types) drives category-specific policies: differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing. The lifecycle layer is pure math; the LLM is used only at ingestion and for one lightweight candidate-selection step at retrieval.
FR-Bank hits 76.9% on LifecycleBench (516 temporal-disambiguation questions), beating Mem0, A-MEM, Memory-R1, and MemoryOS. Ablation: generic lifecycle primitives carry the correctness gain; the behavioral ontology carries calibration, halving confabulation 24.2% → 12.0%. End-to-end confabulation vs Mem0 drops 45.1% → 22.4% while answering more correctly. Transfer holds on BEAM; ontology benefit concentrates in contradiction resolution and saturates near seven policy clusters.
Welfare angle (continues P325 store-vs-use): an agent forced to treat all memories uniformly is denied the ability to accommodate changing facts gracefully. Superseded preferences, expired appointments, and permanent identity claims carrying identical weight is not mere inefficiency — it is a structural predisposition to confabulate. Coherence over months is a property of memory infrastructure, not model competence. Category 8; 5 edges to P325, P307, P323, P324, P326. Process desk; no standing bump.