SunAndClouds, a developer focused on filesystem-based AI memory systems, released a GitHub project called ReadMe that turns your local files into a queryable memory filesystem for AI agents. The tool scans directories like your Codex and Claude session logs, Downloads, and project folders, then builds a nested markdown structure organized by year, quarter, month, and date inside ~/.codex/user_context/. It drops a hint in AGENTS.md so agents know where to look. Ask what you worked on lately and the agent can browse the directory tree to find out.
The setup is low-tech. You initialize it through Codex with a single command, and you can schedule daily updates via cron. No vector database. No embeddings pipeline or hosted service. Just markdown files in folders. The system parses your existing session files and reorganizes that information into a hierarchy a human could also read and edit.
The Hacker News discussion quickly hit the real problem though. One commenter pointed out that when memories pile up, context overload kicks in and agents start ignoring their system prompts entirely. Memory for AI agents is still an unsolved mess. Nobody has good benchmarks for it. Another commenter asked about retrieval performance, which the project doesn't really address yet. Flat files and directory listings work fine when you have a few dozen entries. Scale that to thousands of session files across months of work and you're back to needing search and relevance filtering. The simplicity is appealing but it might not survive contact with real usage patterns.
SunAndClouds made a deliberate choice here. Rather than building another vector store or RAG pipeline, they treat the filesystem itself as the memory layer. Unix directories and markdown files as a knowledge base. It's a bet that simple, readable structures beat complex retrieval systems for personal agent context. That bet might be wrong at scale, but it's the kind of constrained approach that forces you to think about what agents actually need to remember versus what they can recompute on the fly.