Fortunate Recall: Ontology-Driven LLM Memory Lifecycle

Background and Context

Large language model memory systems suffer from a structural flaw that limits their reliability over time. Existing architectures treat every personal fact the user provides as equally important, without distinguishing fact categories or accounting for temporal validity. Memory stores accumulate without bound, and retrieval precision degrades as the corpus grows. The authors identify lifecycle management as the core problem: deciding which memories persist long-term, which should be replaced, at what rate, and conditioned on each fact's behavioral type.

The paper introduces Fortunate Recall (FR), a composable policy layer that addresses this gap. FR classifies personal facts into a behavioral ontology of ten plus one categories, distinguishing facts that hold continuously from those valid only within specific event windows. Each category receives tailored lifecycle strategies rather than a single universal rule. This shifts memory from a passive storage container toward a self-regulating system whose retention and淘汰 decisions rest on fact semantics rather than chronological order alone.

Deep Analysis

FR applies four category-specific lifecycle strategies, all implemented as deterministic functions over metadata functions extracted by the LLM. Differential time decay lets some facts fade slowly while others expire rapidly once their event window closes. Slot-key replacement allows new facts to automatically overwrite older records occupying the same semantic slot, preventing accumulation within a single slot. Event-time validity binds certain facts to explicit validity intervals so they are naturally discarded after expiration. Category-aware retrieval routing directs queries toward the most relevant memory clusters based on category matching.

The infrastructure-agnostic implementation, FR-Bank, lets these strategies deploy across storage backends. On the newly constructed LifecycleBench, a 516-question temporal disambiguation benchmark evaluating judgment over time conflicts and fact updates, FR-Bank achieves a 76.9% pass rate. It outperforms Mem0, A-MEM, Memory-R1, and MemoryOS, which score between 61% and 70.5%. On LongMemEval-S under the standard Wu et al. adjudication protocol, FR-Bank reaches 75.2%, indicating lifecycle strategies impose negligible cost on standard retrieval.

A preregistered ablation precisely locates the gain sources. Replacing the type-aware layer with three generic lifecycle primitives produced no statistically significant accuracy change of negative 1.7 percentage points, with a 95% confidence interval of negative 6.0 to positive 2.7. Generic lifecycle metadata carries the correctness advantage, while the behavioral ontology primarily provides calibration, halving downstream hallucination from 24.2% to 12.0% with p below 0.001. End-to-end, FR-Bank cuts Mem0's hallucination rate from 45.1% to 22.4% for answered queries and from 32.2% to 13.0% for all queries, while raising the proportion of correctly answered questions from 18.6% to 31.2%.

Industry Impact

The paper offers a transferable architectural paradigm for long-term memory systems, reframing memory management as active lifecycle governance rather than passive storage. The infrastructure-agnostic FR-Bank design lowers industrial adoption barriers by integrating seamlessly across storage backends. Decoupling the behavioral ontology from lifecycle strategies lets researchers optimize correctness and calibration independently, providing a clear direction for follow-up work.

The authors open-source the ontology, benchmarks, and code. The newly constructed LifecycleBench and BEAM provide standardized evaluation tools for temporal disambiguation, a long-neglected problem, potentially driving community attention toward memory timeliness. Halving hallucination from Mem0's 45.1% to 22.4% points directly to a key path for improving reliability in real-world deployments.

Outlook

The ranking reproduces on open-weights Kimi K2.5, validating the approach across model families. On the independently built BEAM benchmark, the decomposition approach shows transferability, with 46.8% correctness across 280 questions versus 32.9% for Mem0.

Ontology gains concentrate on contradiction resolution and saturate as the system approaches six strategy clusters. These results suggest lifecycle management will become a standard component of memory architectures, with the behavioral ontology providing calibration benefits that extend beyond raw accuracy gains toward measurable reliability improvements in production systems.