Agent memory architecture is still unsolved
The 'Agent Memory as a File Format' thread gets a blunt response: 'That's a whole lot of text to say it's markdown.' But beneath the snark is a real debate. One commenter building an 'agent fleet' independently landed on a markdown wiki with simple querying as their memory system, calling it 'decently effective.' Others push back that unstructured memory files are the wrong architecture entirely.
This is one of the most practical unsolved problems in production AI systems right now. Agents need persistent memory to be useful across sessions, but every proposed solution, vector databases, structured files, markdown wikis, has real tradeoffs in query reliability, write contention, and context window management.
The diffusion language model thread adds a technical adjacent point: people are actively exploring non-autoregressive architectures that change how models process and revise output. If the underlying architecture shifts, the memory and context management problem looks different too.
So what?
If you are building an agent-based product, your memory architecture is a first-class product decision, not an implementation detail. The current state of the art is embarrassingly lo-fi, markdown files and simple vector search, which means there is real product opportunity for anyone who can make agent memory reliable and queryable at scale without requiring the agent itself to manage its own state.