Governed memory & the graph

Available today

Governed memory is not a chat log and not a vector blob you hope for the best on. It is a structured, cited record of what your team has decided and knows, built so an agent can consume it and a human can audit it.

What is stored

• Documents and notes. Specs, decisions, and design notes, captured as evidence and indexed so they are retrievable.
• Claims. The individual assertions inside those documents (“the rate limit is 100 requests per minute”), tracked at a finer grain than the document so they can be superseded, contradicted, or refined independently.
• Rules. Confirmed constraints that can be injected and, in some cases, enforced.
• Relationships. Typed edges between claims and documents. A newer decision can SUPERSEDES an older one; two claims can CONTRADICTS each other; one can REFINES or REFERENCES another.

Those typed relationships are what make the memory a graph rather than a pile. They are how the system knows that a 2026 decision replaced a 2025 one, and can serve the current answer instead of the stale one.

Evidence is not truth

Everything above can enter as evidence. Evidence is captured and searchable. Evidence is not automatically trusted. A separate step, always involving a human, promotes evidence into trusted knowledge that the system will serve and ground answers on.

Note

This split (capture generously, trust conservatively) is the core of the product. It is important enough to have its own page: the trust model.

Everything traces to a source

Because the memory is built from cited claims, every trusted answer can point at where it came from. When Meetless Agent tells the agent “the rate limit is 100 per minute,” it can name the document and the decision that established it. That is what makes the output checkable instead of merely plausible, and it is the same property that lets the audit trail answer “who decided this.”

How the graph is used

• Injection reads the confirmed, trusted slice and places it in the agent’s context (Injection).
• Ask retrieves over the graph and answers with citations, following SUPERSEDES edges so it returns the current claim, not a superseded one (Ask).
• Capture adds new evidence and proposes new relationships, all born pending (Capture).

The next page explains the rule that governs all of it. Continue to the trust model.