Source: arXiv — 2026-08-05
Summary
"Caching for the Future" operationalizes a finding from animal cognition — that western scrub jays remember not just what and where they cached food, but also how long ago, and adjust retrieval based on how perishable each item is — into an LLM agent memory system called ScrubJay-MEM. Each stored memory is a What-Where-When tuple with a learned perishability estimate and utility horizon, revised at O(1) LLM calls per update, aimed at agents that persist across sessions and otherwise treat all memories as equally durable regardless of how fast their content goes stale.
Key Takeaways
- The core problem: existing agent memory stores treat a fact learned yesterday the same as one learned six months ago, so retrieval keeps surfacing outdated information alongside current facts with no built-in sense of "this has probably expired."
- ScrubJay-MEM's fix is type-conditioned decay — each memory gets an auto-classified perishability coefficient and utility horizon based on its content type, not a single global time-decay applied uniformly to everything.
- Memory updates are retroactive but cheap: revising a memory's decay estimate as new information arrives costs O(1) LLM calls, not a full re-scan of the memory store.
- The paper introduces a new benchmark, the Temporal Generalization Test (TGT), with a Generalization Gap (GenGap) metric measuring how well a memory system holds up on retention intervals it wasn't tuned for — ScrubJay-MEM is reported as the only retrieval-based system with a substantially positive GenGap.
- On MemoryAgentBench's EventQA-64k task, ScrubJay-MEM improves F1 by +2.66 over Mem0 and +3.09 over a Qwen3-Embedding-4B retrieval baseline.
Reel Script
Hook (17s, ~38 words) A bird that hides food for winter knows exactly which stashes have already rotted and which haven't — and most AI agent memory systems don't have that instinct at all. They just remember everything as equally true, forever.
Core Concept (70s, ~155 words) Western scrub jays cache food and later retrieve it, but the interesting part is what they track: not just what they hid and where, but when — and they use that timestamp to skip caches of perishable food that's already gone bad, while still checking caches of food that keeps. That's the "What-Where-When" structure this paper borrows directly. Most LLM agent memory systems store facts with maybe a single global decay curve, if any decay at all — a fact from month one and a fact from yesterday get treated the same at retrieval time. ScrubJay-MEM instead tags every memory with its own perishability estimate and a utility horizon, learned from what kind of memory it is. A note that "the user prefers dark mode" decays slowly. A note that "the deploy is currently broken" should decay fast. The system updates these estimates retroactively as new evidence comes in, and does it cheaply — a constant number of LLM calls per update, not a full memory re-scan.
Hands-On (55s, ~130 words) To actually test whether decay estimates generalize, the authors built a new benchmark called the Temporal Generalization Test, which deliberately holds out retention intervals the system wasn't tuned on, scored with a metric they call GenGap — basically, does your memory system still work when time gaps look different from training. Their headline claim: ScrubJay-MEM is the only retrieval-based memory system tested with a substantially positive GenGap, meaning it doesn't just memorize "how stale things usually are" for one specific time window. On a more standard benchmark, MemoryAgentBench's EventQA-64k task, it beats Mem0 by 2.66 F1 points and beats a Qwen3-Embedding-4B retrieval baseline by 3.09 points.
Takeaway (22s, ~48 words) Most "agent memory" products today are really just vector search with a timestamp bolted on. This paper is a reminder that memory needs a notion of relevance decay baked into its architecture, not tacked on after. If you're building persistent agents, ask what your memory store forgets — and when.