AI Agents Need Docs, Not Memory Plugins
Original: Agents don't need memory, they need documentation
Why This Matters
Challenges a whole category of dev tooling built on assumptions that may be wrong.
Developer Kevin Liao argues that AI memory plugins—built on RAG snippet retrieval—fundamentally misidentify the problem. Agents don't need recall; they need structured, current documentation.
Kevin Liao's essay makes a sharp case: every agent memory plugin on the market is essentially the same product. Transcripts get chunked into snippets, stuffed into a vector database, and the top-5 most similar hits get injected into each prompt. Fancier variants add multi-tier classification, background "dreamer" daemons that rewrite memories overnight, context compression, and rerankers—but the underlying architecture is identical, and identically flawed.
The core problems Liao identifies: similarity search can't distinguish current facts from stale ones; RAG snippets strip away motivation and context; and agents can't search for knowledge gaps they don't know exist. Worse, with thousands of embeddings in a database, there's no practical way to audit which memories are wrong, outdated, or silently warping the agent's behavior.
His counterargument is blunt—no engineer rewinds a three-year-old meeting to recall a feature constraint. They check written records. The fix isn't smarter recall; it's documentation that agents can actually read and act on. Liao points to AGENTS.md-style files as the starting point: structured, human-readable, maintainable context that survives codebase churn. He frames this as increasingly urgent precisely because AI is now shipping code faster than anyone reads it.