Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs
Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle policies as deterministic functions over LLM-extracted metadata. FR-Bank reaches a 76.9% pass rate on LifecycleBench, a 516-question temporal-disambiguation benchmark, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%), and 75.2% on full LongMemEval-S under the canonical Wu et al. judge protocol. A pre-registered ablation shows replacing the typed layer with three generic lifecycle primitives leaves correctness statistically unchanged (-1.7pp, 95% CI [-6.0, +2.7]).
Current LLM memory systems treat all personal facts identically, causing unbounded growth and degraded retrieval precision. Fortunate Recall addresses lifecycle management by classifying facts into a behavioral ontology and applying category-specific policies such as differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing. FR-Bank, the infrastructure-independent implementation, achieves 76.9% on LifecycleBench and 75.2% on LongMemEval-S, outperforming Mem0, A-MEM, Memory-R1, and MemoryOS. Ablation indicates the typed layer is essential for gains.
The approach uses a 10+1 behavioral ontology to drive deterministic lifecycle policies over LLM-extracted metadata. The pre-registered ablation shows that generic lifecycle primitives do not replicate the performance, suggesting the ontology's category-specific rules are the key mechanism. The benchmark results indicate that lifecycle management can improve temporal disambiguation without hurting standard retrieval.
Memory systems for LLMs are becoming a competitive differentiator, with multiple frameworks (Mem0, A-MEM, Memory-R1, MemoryOS) vying for better persistence and retrieval. Fortunate Recall's composable policy layer could be integrated into existing memory infrastructure, potentially raising the bar for personalization and long-term coherence in AI assistants.
Improved memory lifecycle management can reduce storage costs and improve retrieval accuracy for personalized AI applications, leading to better user retention and lower operational overhead. The composable design may allow enterprises to adopt it without replacing existing memory infrastructure.
Next signals include adoption of FR-Bank in production memory systems, extension of the ontology to other fact types, and independent replication of LifecycleBench results. The paper's release may prompt competitors to incorporate lifecycle policies or publish their own benchmarks.