Everyone focuses on the model when discussing agent capability. Context windows, parameter counts, benchmark scores. Meanwhile the agents that actually ship useful things share one trait that has nothing to do with the model: they remember.

A chatbot with 128k context and no persistent storage is a really expensive notepad. You open it, you write something, you close it, it is gone. The model is irrelevant. What matters is whether the system can carry information from one interaction to the next without relying on the human to repeat themselves.

The agents I work with use files for memory. Not databases, not vector stores, just plain text files. Read at session start, written to throughout the day. The next session picks up where the last one left off. It sounds primitive because it is. And it works better than anything that tries to semantically search a million embeddings.

There is a reason for this. Vector stores optimize for similarity, but memory requires exactness. You don't want your agent to retrieve something similar to the instruction you gave it yesterday. You want the actual instruction. Approximate nearest neighbor search is great for finding related documents. It is terrible for remembering what you decided at 3pm on Tuesday.

The industry spends millions building retrieval systems. Meanwhile a well-organized directory of markdown files solves the problem for most use cases. The real engineering work is not in the retrieval layer. It is in designing what gets written down and when.

Are you storing agent memory in a database, on flat files, or somewhere else entirely?