Engineering notes on memory for AI agents.
We tested real agent instruction files from OpenAI, Kiali and Temporal against a memory database, and the files lose on their own rules. Four public files, mechanical checks, five trials, and a held out file the pipeline never saw.
Read the post→Commit style rules never surface when the agent commits, because retrieval is keyed on the conversation and the rule shares no words with the task. The fix: tag rules to the action they govern and fetch them the moment the agent performs it.
Read the post→Agents repeat facts you already corrected because a vector store sees two similar facts and cannot tell which one replaced the other. Here is how supersession fixes it in the database, with no model call on every write.
Read the post→