What my agents are allowed to remember

Every agent platform now includes memory, enabling persistence across sessions. Initially, this feature appeared to settle questions, but disappointingly, agents often retained irrelevant information. The source reveals that memory was organized like a library and accessed like a headline. The issue stemmed from the memory store's two-layer structure.

The index file, loaded into every session, consisted of one line per memory, while the body file, opened only when an agent needed to read it, contained the actual facts. The titles provided by the index were insufficient to inform agents mid-task about the importance of the associated body files. Consequently, numerous valuable memories became redundant and unused, leading to the need for a routing rule to address the problem.

This one-line rule entailed placing actionable atoms in the index line itself and reserving the body for worthier, deliberate fetches, including the 'why' and the full runbook. The index line should be written as if it were the only information the agent would ever see, as it typically is. In addition to the technical issues, the source highlights the importance of admission.

Memories should pass three tests to be admitted: not derivable, durable beyond the conversation, and behavior-changing. If these criteria are not met, the memory is rejected. The source also discusses decay, a problem that often goes unnoticed. Stored claims are assertions about the past and should be verified on repeat or deleted if outdated.

The author emphasizes the significance of maintaining a lean store, where every entry undergoes a three-test evaluation to ensure its relevance and value. The author concludes by stating that after implementing these changes, the store effectively removed unnecessary noise, and agents wasted less time on redundant information.