Intel / article

Agents don't need memory, they need documentation

Source: liao.gg
Author: kmeh
Date: Published Oct 3, 2026

From the Daily Brief

The case against agent memory plugins in this essay rests on one mechanism. Retrieval by similarity returns whatever text looks closest to the query, and an old, superseded note about a function can look just as close as the current one. On a codebase that changes every day, that means an agent can confidently recall a fact that stopped being true last week. The author argues that this is a structural limit, not a tuning problem, and that the better fix is document-based context: written, maintained documentation that is updated when the code changes and read as a whole. The argument lines up with how many teams already work, keeping agent instructions in checked-in files rather than in an opaque store. It is an opinion piece rather than a measured study, so it does not quantify how often stale recall causes errors. Teams running memory plugins can test the claim directly by checking how their tool handles a fact that has been revised.

Read the 2026-10-05 Brief →

Context

Operator Memory is a plugin intended to help coding agents carry project knowledge between sessions through editable documents. Its creator, Kevin Liao, describes loading a small catalog at session start that explains what the documents cover and when to consult them. The opens relevant material before working, including requirements and the reasoning behind earlier decisions.

The agent is instructed to update those documents while the task is still in , replacing outdated facts. Liao says he still needs to request revisions or trimming. Because the records are ordinary text files, developers can inspect and correct the project knowledge that future sessions will use. His article makes that maintenance part of the workflow.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
RAG — Retrieval-augmented generation — fetching relevant documents first and pasting them into the model's context so it answers from your data instead of memory.

Read the source liao.gg