Docs/Learn

Three verdicts

A plain memory can only back a claim up. A grounded one is also turned into subject → relation → object facts that a Z3 solver can reason over, which is what makes refuted possible: a wrong number becomes a visible failure instead of a quiet guess.

What to ground

Ground a fact

remember({content: "The HNSW index is always reconstructable from the embeddings table.",domain: "architecture",tags: ["arch", "constraint"],priority: 10,namespace: "my-project"})

It's stored as a normal memory and promoted to solver facts straight away. For facts you're less sure of, pass promote_immediately: false and let them earn their place through repeat confirmation.

Ground your docs

Pass a domain when you index a folder of Markdown — ADRs, design principles, architecture notes. Only prose is grounded; code bodies never are.

code({action: "ingest",path: "./docs",domain: "architecture",ground_immediately: true,namespace: "my-project"})

Check claims

verify splits text into claims and judges each one. Agents are reminded to run it before stating facts; you can call it any time.

verify({text: "The HNSW index persists across restarts. It is rebuilt from the embeddings table.",domain: "architecture"})// claim 1 → refuted// claim 2 → grounded

For one precise relation — "is A a B?", "does A call B?" — prove returns proved, refuted or unknown with a certificate of the facts it used.

prove({ claim: "HNSW index is durable state", domain: "architecture" })// → { verdict: "refuted", certificate: { axioms_used, proof_steps } }

Manage domains

domain({ action: "load", source: "./domains/architecture.domain" })domain({ action: "status", name: "architecture" })   // fact count, modedomain({ action: "export", name: "architecture" })   // back to a file, for git

Something missing or wrong? Email support@memorylayer.in or open an issue.