PreCoMem is a memory‑consolidation framework inspired by predictive coding that represents dialogue memory as confidence‑weighted beliefs about the user’s state. Its core signal, Effective Surprisal, combines semantic deviation and contradiction evidence, and is down‑weighted when retrieval is diffuse and unreliable.
Based on Effective Surprisal, a two‑threshold gate maps each turn to three deterministic updates: MAINTAIN reinforces supported beliefs, PROFILE stores ambiguous cues as low‑confidence hypotheses (acting as a waiting room), and CORRECT softly decays reliably contradicted facts.
The PROFILE stage prevents premature commitment to noisy evidence, allowing the system to keep memories stable while still adapting to genuine user changes.
Experiments on LoCoMo, LongMemEval‑S, and PersonaMem‑v2 show consistent state‑of‑the‑art accuracy. The newly built DiMoBench further confirms that PreCoMem adapts to non‑stationary belief drift without over‑reacting to noise.
The code is released at https://github.com/ZhongWishing/PreCoMem.
Review