An AI agent's memory has to change — but so do its mistakes

The main change in MindMemOS is that an AI agent's memory is treated here not as an archive but as a system that changes its own structure and rules of operation. That increases autonomy while at the same time making errors more dangerous: a wrong memory may not merely be stored but built into the agent's subsequent behaviour.

The project organises open data through a single structure of entities, properties and time. MindMemEvolve then fits a memory schema to the specific scenario, and a dreaming mechanism merges duplicates and tries to resolve contradictions. In other words, the agent should not only remember more but periodically put its knowledge in order.

There is a second layer too — MindSkillEvolve. It turns task-execution trajectories into reusable skills and gradually improves them. On the reported tests that yielded a gain of 9.2 percentage points in SpreadsheetBench success relative to the initial skill set. On LOCOMO the system showed 94.03% accuracy, on PersonaMem 70.63%.

But autonomous memory evolution does not abolish oversight — it makes it more important. For that MindMemOS uses implicit corrective feedback from a human: it helps find memories that are potentially inaccurate or a poor fit for the task. At the same time it is unclear from the description how often such errors arise, how much constant memory rebuilding costs, or how the system behaves beyond the named datasets.

The practical conclusion is simple: an approach like this is useful where an agent works in one domain for a long time and accumulates repeating experience. But launching self-editing memory without checks is not an option — otherwise we get not a smarter assistant but a system confidently learning from its own misconceptions.

Are you prepared to let an agent change the structure of its own memory when checking for errors still requires a human? Source: cs.AI updates on arXiv.org