Abstract

Memory systems for interactive agents store observations — facts about the world and the user. We argue the system's own reasoning traces ("thoughts") require a structurally different store: a separate spectrum with its own much faster decay law, and a gated bridge to long-term memory that only a confirmed outcome can open. A thought's retention follows an exponential decay with a roughly 4-hour half-life; cold thoughts are pruned. A thought is crystallized into a durable fact only when a later, externally-produced outcome confirms it — unconfirmed reasoning is never allowed to become long-term "truth," which makes the gate an anti-hallucination mechanism by construction (the system cannot promote its own beliefs, only outcome-confirmed ones). We describe the deployed implementation across two machines feeding one remote store (17,150 total memories; 6,908 captured thoughts; 107 crystallized) and report honestly which mechanisms have and have not yet fired in production, including a defect: the re-reference count has never incremented in deployment. We propose a pre-registered experiment comparing gated crystallization against keep-all, discard-all, and heat-only promotion. Finally we show the spectrum's second use, which we believe is underappreciated: a persistently captured thought stream is an external honesty audit — divergence between what a system thought and what it said is detectable by a party the system cannot edit. An initial prior-art search (Generative Agents, MemGPT, ACT-R, chain-of-thought logging, process supervision, Reflexion, Voyager) finds each ingredient occupied but the conjunction — own-trace fast-decay spectrum plus independent-outcome gating plus external audit — separable, with Voyager's execution-verified skill library the nearest neighbor.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS