What ThinkingRoot is

ThinkingRoot is a cognition database — a database whose unit is not a row but a living memory. You compile your sources (docs, code, conversations, data) into a verifiable cognition graph, and then you recall, ask, and reason over it with grounded, cited answers that never hallucinate. It is the memory and reasoning layer for AI applications and agents.

Where a normal database stores and returns rows, ThinkingRoot compiles knowledge into atomic claims, links them into a graph, tracks where every fact came from, and serves them back as grounded context — to your LLM or to ours. The result: answers that carry citations, say "I don't have that in memory" instead of making things up, and let you delete a memory cleanly because every claim has provenance.

The eight verbs

ThinkingRoot is a living Mind, not a passive store. Everything it does is one of eight verbs:

• remember — write a fact into memory, with provenance.
• recall — semantically retrieve the relevant memories for a query.
• learn — improve which memories surface, from which ones actually get used.
• sleep — consolidate and age memory in the background.
• dream — synthesize new insights from existing memories, quarantined until verified.
• predict — make grounded "what happens next" predictions, or refuse if unsupported.
• fork / merge — branch memory like git: explore, then merge what verifies.
• inherit — pass a memory's knowledge forward (capsules, brain packs).

How it works, end to end

1. Compile. You add sources to a workspace. The engine extracts claims (atomic, deduplicated facts), links entities, attaches provenance (which source, which bytes), and supersedes stale facts. This is the "compile" step — it turns raw documents into a queryable cognition graph.
2. Recall. A query is embedded, matched against the graph (vector + graph traversal), and reranked. You get the most relevant claims, each cited.
3. Answer. Those claims become grounded context. An LLM (yours via MCP/SDK, or ours) synthesizes a cited answer — and abstains when the claims don't support one.

Because retrieval returns compiled, deduplicated, cited claims instead of raw document chunks, the context handed to any LLM is far smaller (often ~7× fewer tokens) and provably grounded.

What you build on it

• AI apps with real memory — chatbots, copilots, and agents that remember across sessions and answer from your knowledge with citations.
• Root Functions — durable compute that lives inside the brain (see Root Functions).
• Self-extending agents — capabilities that are authored and verified at runtime.
• Multi-tenant products — one workspace per project, per-user sub-isolation, all enforced at the engine boundary.

OSS engine + managed cloud

The engine is open source — the root CLI, root serve (a self-hostable database), the desktop app, and the SDKs. ThinkingRoot DB Cloud is the managed, multi-tenant platform: accounts, billing, an authenticating gateway, and a provisioner that runs one engine per project. The cloud and local use the same engine, so anything you build or learn locally transfers directly to the cloud.