Agent Memory

Agent memory is state an assistant keeps across conversations, since a Context Window does not survive a session. In its simplest and most durable form it is a set of small files, each holding one fact, with enough metadata to decide later whether the fact is relevant.

What makes memory useful is discipline about what belongs in it. Things worth storing are those a future reader could not reconstruct: a preference and the reason for it, a decision and what it foreclosed, a constraint discovered the expensive way. Things not worth storing are those the artifact already records — the structure of the code, the history of past fixes, anything already written down elsewhere. Duplicating those creates two sources that drift; see Documentation Rot.

Memory is also dated. A stored fact reflects what was true when written, and a note naming a file, a flag, or an interface can be silently wrong months later. Treating recalled memory as a hypothesis to verify rather than a fact to act on is the habit that keeps it from becoming a source of confident errors.

The related design question is retrieval: a memory that is never surfaced at the right moment is inert, and one that is always surfaced is just a longer prompt. Summaries carrying enough to judge relevance, with the full text loaded on demand, is the shape that scales.