ThinkFlow is an end‑to‑end latent memory framework designed for lifelong conversational agents. By dynamically compressing dialogue flows into probabilistic latent memory skills at test time, it bypasses the information bottleneck of explicit textual memories and encodes complex user states into disentangled continuous vectors, eliminating semantic interference.
To address cold‑start and continuous personalization, the framework adopts a test‑time evolution paradigm. An initial state is bootstrapped via teacher‑guided latent alignment, after which a self‑supervised next‑utterance prediction task continuously refines the memory vectors, enabling label‑free lifelong learning.
Extensive experiments on long‑term conversation benchmarks show that ThinkFlow markedly outperforms existing explicit textual memory systems, delivering more personalized and context‑accurate responses across multi‑session interactions.
Review