Agent memory should model the user’s world, not just store every chat
A common way to give an AI is to save every user message, turn it into , and later retrieve similar passages for the model’s context. This is essentially a searchable chat archive, so it is easy to build and looks effective in a demo. A year of firsthand development found problems that better search alone cannot solve.
When information recorded in March is corrected in June, both versions remain available and the model may choose the outdated one. References such as “he,” “that client,” and “the project” are not reliably connected to the same real person or project across conversations. A commitment in one conversation and its in another also appear to be unrelated records.
Because these relationships were never stored, better or reranking cannot reconstruct them. A stronger ls the user’s world and stores people, projects, and commitments as distinct entities instead of loose sentences.
Key points
- search can return an outdated fact alongside its newer correction.
- Names, pronouns, and labels across conversations must be linked to the same real entity.
- Related events, such as making and completing a commitment, need an explicit connection.
- Better search cannot recover relationships that were never stored.
- Store people, projects, and commitments as distinct entities rather than scattered sentences.