Long conversational agents are now ubiquitous, yet they must retain distant dialogue turns so users do not have to repeat instructions or context. The main difficulty is that instructions and context evolve, requiring a memory that preserves both current and historical states, distinguishes stale from active knowledge, retrieves evidence relevant to a query, and avoids repeatedly invoking a large language model to rewrite past interactions.
We introduce MINDSET, a memory controller that stores conversations as immutable episodes and organizes them into versioned schemas via minimum‑energy state transitions. Each incoming episode may reinforce, supersede, split, or create a schema. The transition decision balances representation distortion, contradiction, historical damage, fragmentation, and internal inconsistency, while a hysteresis mechanism prevents isolated contradictions from prematurely overwriting stable memory.
Evaluated on a reproducible set of 850 questions (700 LoCoMo + 150 MemoryAgentBench) against five memory systems, MINDSET achieves the highest LoCoMo answer F1 and markedly improves retrieval ranking metrics (Recall@8, MRR, nDCG@8) over the second‑best method LightMem (p < 0.05).
Review