Files, not hidden state
Durable memory is inspectable, diffable, portable, and deletable by the person who owns it.
For OpenAI-adjacent builders and agent developers
A local-first memory pattern for AI agents.
PersonalOS keeps durable context in editable Markdown files, with agent-agnostic policy and runtime adapters instead of hidden product memory.
Conversation → proposed memory → user review → Markdown files → future agent context.
Policy plus adapters
AGENTS.md holds canonical policy. Runtime files such as Claude, Codex, and Agy adapters stay thin.
Consent before durability
Agents can propose what to remember, but durable personalization should be reviewed before it is saved.
Why it matters
AI memory is becoming infrastructure.
If assistants remember more, builders need patterns that let users inspect and move that context. PersonalOS explores one practical local-first answer.
Experiment
Fork the repo, inspect the file layout, and try new memory or skill conventions.
Integrate
Use the public policy/adapters model as inspiration for other agent runtimes.
Contribute
Improve upstream docs, scripts, and templates without exposing private memory.