The mental model
Why ThinkingRoot is a database of grounded claims, not a pile of text or vectors.
Most "knowledge" systems store text and hope a model retrieves the right chunk. ThinkingRoot stores claims: atomic, typed statements, each locked to the exact source span it came from. The graph of those claims — plus the entities they mention and the relations between them — is what your AI reads. That's why answers can cite their evidence and why the engine can refuse to answer when the evidence isn't there.
The chain of custody
source → compile → witnesses + claims → entities + relations → graph
│
ask / search ◀─────────────┘
(answers cite the claims)- A source is registered (a file, document, or text), content-addressed so changes are detectable.
- Compilation reads the source and derives witnesses (content-addressed units of primary bytes) and claims (typed statements with a confidence score).
- Claims mention entities (people, systems, concepts) connected by typed relations.
- When you ask, the engine retrieves the relevant claims and builds the answer from them — citing each one.
The terms, in one line each
| Term | What it is |
|---|---|
| Source | The origin of knowledge — content-addressed by BLAKE3 for change detection. |
| Compilation | Extracting claims/witnesses from sources and admitting them to the graph. |
| Claim | An atomic, source-locked, typed, timestamped statement of fact (with a confidence score). |
| Witness | A content-addressed unit derived from primary bytes via a named rule; dedupes across workspaces. |
| Entity | A named thing — person, system, concept, file, function — keyed by canonical name. |
| Relation | A typed, directed edge between two entities, backed by the claims that evidence it. |
| Branch | An isolated, copy-on-write fork of the graph; sessions get their own auto-merging branch. |
| Health | The result of verifying the graph — contradictions, stale claims, grounding failures. |
Why this matters for what you build
Because every claim is locked to a source span, an answer isn't "the model's opinion" — it's a set of citations you can click through and audit. And because the graph knows what it contains, "I don't have that" is a first-class answer.
Read on for each piece in depth.