Problem
Research questions and friction points this paper is trying to address.
LLM agent memory
temporal decay
episodic memory
memory validity
outdated facts
Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems
🤖 AI Summary
Current memory systems in large language model (LLM) agents struggle to differentiate the temporal relevance of information, leading to retrieval contaminated by outdated content. Inspired by the episodic memory mechanisms of western scrub-jays, this work proposes a novel memory architecture that represents memories as What-Where-When triplets and assigns each memory type a learnable, type-conditioned temporal decay coefficient. Coupled with query-adaptive scoring, the system enables efficient retrieval and O(1)-complexity retrospective updates. This is the first effort to integrate the biologically observed principle of type-dependent temporal decay from episodic memory into artificial agent systems. The authors also introduce the Temporal Generalization Test (TGT) benchmark for evaluation. Experiments demonstrate that the proposed method achieves a +0.108 generalization gain on TGT and outperforms Mem0 and Qwen3-Embedding-4B by 2.66 and 3.09 F1 points, respectively, on the MemoryAgentBench EventQA-64k task.
📝 Abstract
LLM agents that persist across sessions accumulate stored memories whose validity varies enormously by content type, yet existing memory architectures treat all memories as equally persistent and systematically contaminate retrieved context with outdated facts. We show that per-memory, type-conditioned temporal decay, a property of western scrub jay episodic memory, can be operationalized as an auto-classified coefficient $π_i$ in an external LLM-agent memory store, yielding ScrubJay-MEM: each memory is encoded as a jointly-bound What--Where--When tuple with an estimated perishability $π_i$ and utility horizon $τ_i$, retrieved by query-adaptive scoring, and revised retroactively at $O(1)$ LLM calls per update. We introduce the Temporal Generalization Test (TGT), a benchmark with held-out retention intervals and a Generalization Gap (GenGap) metric. On TGT, ScrubJay-MEM is the only retrieval-based system with substantially positive GenGap ($+0.108$); on MemoryAgentBench EventQA-64k it improves F1 by $+2.66$ over Mem0 and $+3.09$ over Qwen3-Embedding-4B under a llm backbone. A decay ablation collapses GenGap by $5.7\times$, establishing type-conditioned decay as necessary for the result. Gains narrow under stronger backbones and reverse on fact-consolidation tasks, scoping the contribution to temporal reasoning over perishable facts.
Innovation
Methods, ideas, or system contributions that make the work stand out.
type-conditioned temporal decay
episodic memory modeling
memory perishability
query-adaptive retrieval
Temporal Generalization Test
🔎 Similar Papers
No similar papers found.
