LLM memory management improves personal fact accuracy and persistence
Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs
Artificial Intelligence
Summary
Language models keep track of many facts about people, but treating all these facts the same way makes the memory large and less accurate. The authors created Fortunate Recall, a method that sorts facts into different types and manages how long to remember each based on those types. This helps the model keep important information longer and forget less useful details sooner, improving accuracy and reducing made-up answers. Their system outperformed other methods on several tests and cuts wrong answers nearly in half while answering more questions correctly.
What this means in practice
- •For conversational ai developers: Improve chatbot memory by selectively remembering user facts with tailored decay rates to reduce errors and enhance response accuracy.
- •For personal assistant platforms: Manage user data lifecycle to maintain relevant information longer while removing outdated or less useful facts to improve personalization quality.$Commercial implications: Enables smarter personalized assistants that reliably track user information over time, boosting user satisfaction and trust.
Authors
Ansuman Mullick, Eray Tüzün
Abstract
Current LLM memory systems treat all personal facts identically, so stores grow without bound while retrieval precision degrades. The core challenge is lifecycle management: which memories should persist, which should be replaced, and at what rate, conditioned on the behavioral type of each fact. Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle policies (differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing) as deterministic functions over LLM-extracted metadata. FR-Bank, our infrastructure-independent implementation, reaches a 76.9% pass rate on LifecycleBench, a new 516-question temporal-disambiguation benchmark, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%), and 75.2% on the full LongMemEval-S under the canonical Wu et al. judge protocol, so lifecycle policies impose no measurable cost on standard retrieval. A pre-registered ablation locates the gains: replacing the typed layer with three generic lifecycle primitives leaves correctness statistically unchanged (-1.7pp, 95% CI [-6.0, +2.7]), so the generic lifecycle metadata carries the correctness advantage, while the behavioral ontology carries calibration, halving downstream confabulation (12.0% vs 24.2%, p<0.001). End-to-end, FR-Bank cuts confabulation from Mem0's 45.1% to 22.4% over answered queries and from 32.2% to 13.0% over all queries while answering more of them correctly (31.2% vs 18.6%); the ranking replicates on the open-weight Kimi K2.5. The decomposition transfers to BEAM, an independently built benchmark: 46.8% correct vs Mem0's 32.9% over 280 questions, with the ontology's benefit concentrated in contradiction resolution and saturating near seven policy clusters. The ontology, benchmark, and code are released.