Giving an AI agent memory is easy. Keeping it true is the hard part.

Memory in AI agents is simple to implement, but maintaining its accuracy is challenging. All that is required is a directory of markdown files containing one fact per file, along with an index that gets loaded at the start of each session. However, even the mechanism itself is not the issue. Despite running such a system for about four months across around forty projects, four separate memory entries were found to be confidently wrong, each steering decisions for weeks.

The problem lies in the inference drawn from the memory entries, which become conclusions without their reasoning left behind. Unlike typical notes, these entries were written carefully at the moment of work, with full context, recording judgments as facts. Once an entry is written, the reasoning is lost, and only the conclusion remains.

This becomes problematic as conclusions do not carry their own expiry date. The solution is to add a re-check command to each entry, making the claim falsifiable and allowing future sessions with no context to catch errors quickly. Additionally, it's crucial to distinguish between observed facts and inferred conclusions, as the latter can become outdated.

When memory and the filesystem disagree, the filesystem should be trusted. The risk with agent memory is not hallucination but decay, where entries become snapshots of moments that do not reflect the current world. Regular verification from outside the system that made the claim is essential, as relying solely on the system's opinion can lead to errors.