A new memory system for AI chatbots decides what to keep, what to update, and what to quietly retire, not by re-reading the whole conversation every time.
Researchers introduced MINDSET, a memory controller for long-running conversational AI agents. Rather than storing a flat transcript or periodically asking an LLM to rewrite a summary, it keeps each exchange as an unchangeable 'episode' and sorts those episodes into versioned schemas, essentially evolving topic folders that can merge, split, or get superseded as new information arrives. A set of rules modeled on minimum-energy transitions decides whether a new episode reinforces an existing schema, replaces it, or forks into a new one, weighing contradiction, fragmentation, and how much history would be lost. The team tested it against five existing memory systems on 850 questions drawn from the LoCoMo and MemoryAgentBench benchmarks.
Most chatbot memory today either hoards everything, which gets slow and expensive, or leans on repeated LLM summarization, which is lossy and compounds errors over time. MINDSET posted the highest answer accuracy on LoCoMo and beat the next-best system, LightMem, on retrieval ranking, with similar gains showing up across different underlying models. That is a real argument for treating memory as structured, versioned state rather than an ever-shrinking summary.
It is one paper beating five baselines on a fixed benchmark, not a shipped product. The real test is whether minimum-energy schema transitions hold up against years of messy, real chat history instead of curated test questions.