What's this about?
Cal Paterson argues most agent memory systems are bad in three distinct ways, and proposes plain markdown files plus search instead. The three failure modes he names: systems that deliberately tie you to a specific harness, usually written by the lab renting you that harness, which mine your conversation history and end up storing memories mostly about *you* when information about the *world* is more useful; systems that are ludicrously complicated, one prominent example needing pgvector, a Neo4j graph database and its own LLM just to decide what is worth remembering; and "High Modernist" schemes that imagine an idealised rationalist memory, inevitably involving a graph and sometimes logical propositions.
Why would complexity make memory work worse rather than better?
Because the model is the consumer. A memory system's job is to put the right text in front of a model that is already good at reading text. Elaborate structure means the model must first understand your schema, which spends capability on the wrong problem — and it fails to scale with the frontier, because next year's model is better at reading prose and no better at your graph. The lab-specific version has a second problem that is commercial rather than technical: it is an attempt to move from the competitive API business into the stickier platform business.
What's the community saying?
- The dismissive top comment — "that's a whole lot of text to say it's markdown" — followed by the fairer correction that it is markdown *plus* retrieval, and that RAG over hyperlinks is the convincing part.
- A resigned observation that gets agreement: it feels like 50% of AI progress is finding different ways to say "tell the model things in English."
- The strongest technical support: agents navigating a codebase don't drill down through plausible directory names, they grep the whole repo for keywords — so structure you impose is largely bypassed anyway.
- A dissent worth keeping: an unstructured pile of memory files may not be the answer either, and "the agents will grep anyway" is a description of current behaviour, not a design principle.
- A practitioner reports the same conclusion independently: a markdown wiki with simple querying works well, and the real unlock is a skill that reduces a session into durable lessons.