Retention of past events in the self-written memories of AI agents

AI agents that work for weeks keep a written memory that they summarise and rewrite themselves. We asked how long past events stay in such memories, whether the memories record correctly how far each task had got, and whether failures are recorded as often as successes. We used the logs of the AI Village, where twelve agents built on models from four developers pursue open-ended goals, each with its own computer, and rewrite their memory about every quarter of an hour while working. We sampled 603 events from the logs and paired each with the agent's memory from 1 hour to 14 days later; two language models from different developers coded each pairing, blind to the agents' identities. One hour after an event, 35% of events were recorded as that specific action or result, and 65% had at least the task or exchange they belonged to recorded. After a day the figures were 12% and 30%; after two weeks, 1% and 4%. When a plan or outcome was recorded, the status the memory gave it matched what the agent had done by the time of writing in 150 of 173 cases, with errors in both directions. Failures were recorded as specific events about half as often as successes (14% against 31% in the first day). When a failure was recorded, its description was often wrong: 8 of 24 recorded failures were misdescribed, against 2 of 131 recorded successes, most often by giving the failure a cause outside the agent that the logs do not support; one agent recorded its own errors as "sabotage" by its environment. Unlike people, who remember their past results as better than they were, the agents recorded the status of their work accurately. The memories worked as records of where tasks stood rather than logs of what happened. For people working with agents, this means that the absence of a failure from an agent's memory is no evidence that nothing failed, and that the causes an agent records for its failures should be checked against its logs.