I build trustworthy memory and terrain for AI agents — so retrieval can find things, and ceremony decides what is allowed to mean something.

AI assistants forget between sessions — and the usual fixes make it worse. Memory files rot. Summaries come back as facts. A model's guess gets written once and every future session inherits it as ground truth. Similarity can surface text. Similarity cannot tell you whether to trust it.

The front door. A manual plus working examples of marbles: small inheritable environments for recurring AI work. Context is labeled by provenance, writes to trusted memory pass through gates that can refuse, and hostile tests prove the refusals are real. Start with START_HERE.md.

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

A constitution and SQLite schema for provenance-aware agent memory. Every entry carries who produced it, what kind of thing it is, and where it came from. Nothing becomes ground without a recorded authority act. Similarity can retrieve. Similarity cannot promote.

Use it if: you're building agent memory or RAG and need to know whether to trust what came back, not only what matched.

A populated, local English Wikipedia hangar: 8.4 million article bearings, offline canopy search, and exact depth only where someone walks. Search returns directions, not evidence. Install with pip install "wiki-forest[semantic]".

Use it if: you want a shared wild substrate agents can search and cite without confusing retrieval with adopted truth.

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 whether it worked.

The_Biodome — experimental residency: a sealed world, a Forest of sources, and time enough to see what kind of inhabitant emerges.