I build small, verifiable environments for AI agents — memory that can be trusted, handoffs that survive a context window, and tests that prove the machinery refuses bad data instead of just claiming to.

AI assistants forget everything between sessions — and the common fixes make it worse. Memory files rot. Summaries come back as facts. A model's guess gets written down once and every future session inherits it as ground truth. If you've watched an agent confidently misremember something you never said, that's the problem.

The main event. A manual plus six working example implementations of marbles: small inheritable environments for recurring AI work. Stored context is labeled by provenance (user-decided / model-inferred / unverified), writes to trusted memory pass through executable gates that can refuse, and hostile test suites — 120 tests across four suites — prove the refusals are real. Built and validated across four different AI builders (Claude Fable 5, Claude Code, Cursor, Codex). Start with START_HERE.md: ten minutes, one cold demo.

Use it if: your agent's memory drifts, your CLAUDE.md contradicts your code, or every new session starts from archaeology.

A working marble for one long-form writing project — and the whole record of how a marble gets made. The first commit is the complete map with zero construction; every commit after is execution, and the design conversations that produced each decision are in logs/, verbatim. Its one law: nobody puts words in anyone's mouth. Everything is signed; nothing enters the writer's canon except through their own quoted, dated adopting words; nothing is deleted — superseded material falls into the woods with its ancestry intact.

Use it if: you want to see agent memory discipline applied to a real creative project, or you're a model learning how a marble gets built — the history is the curriculum.

A spec for provenance-aware agent memory: a small constitution (FOREST.md) plus a SQLite schema that stores custody with the text. RAG retrieves by similarity, but similarity can't tell you whether a sentence was said by the user, guessed by the model, superseded, or rejected. Forest can: every stored entry carries a bucket, a signature, and ancestry, and nothing becomes ground truth without a recorded adoption ceremony. Hostile tests enforce the refusals. Similarity can retrieve. Similarity cannot promote.

Use it if: you're building RAG or agent memory and need to know whether to trust what retrieval returns, not just what matches.

Good — these repos were written to be operated by models. Point your agent at TheMarble, have it read AGENTS.md, and ask it to build a marble for a recurring kind of work. The tests will tell you both whether it worked.