Facts become pages. Links become a graph.
Agents do not write essays into a folder of markdown. They write subject–predicate–object triples: who, what relation, to what. Each subject and each object is a node. Each predicate is an edge. The result is a wiki that assembles itself, one claim at a time.
references and tag coexist; functional ones keep a single current value.
{
"tool": "remember",
"args": {
"subject": "acme-corp",
"predicate": "decision_maker",
"object": "sarah-chen",
"source": "call-2026-06-18",
"confidence": 0.97
}
}
Every claim cites its source.
A wiki is only as good as its citations. In abagraph, provenance is not an afterthought you bolt on — every fact carries where it came from and how sure we are at write time. Click any claim and the trail is right there. Below your threshold, the agent abstains rather than guess.
Source, recorded-at, and valid-from travel with every fact. The agent can verify any claim instead of taking the model's word for it — and so can you, on the entity page.
Writes below your threshold land as Candidates, not truth. Re-asserting at full confidence promotes them. The wiki refuses to launder a guess into a fact.
Every page has a version history.
Wikis keep their edits. So does abagraph — but at the level of meaning. Because storage is bitemporal, nothing is ever overwritten. When a claim changes, the old one is closed as Superseded and the new one opens. Ask as_of(T) and the page reconstructs itself as it stood that day.
{
"tool": "recall_as_of",
"args": {
"subject": "acme-corp",
"as_of": "2026-03-01"
}
}
Valid time is when a fact was true in the world. Transaction time is when you learned it. Tracking both lets you audit not just what changed, but when the agent found out — the difference between a wrong-then-corrected belief and a fact that was always wrong.
No migration, no snapshot job. Time travel is how the store works, not a feature on top of it.
Two wikis, one engine.
The Wiki is the knowledge your agents build — entity pages, claims, and history that grow from what they learn. The product docs are the reference you read — the API, the data model, and the recipes for writing and recalling those facts. Same engine underneath, different reader.
- •An entity page per subject and object
- •Claims with source, confidence, and status
- •Version history and
as_oftime travel
- •REST + MCP API reference
- •The bitemporal data model and conflict rules
- •Recall, walk, and
build_contextrecipes
Give your agents a memory they can browse, cite, and trust.
Entity pages, provenance on every claim, and time travel per fact — a wiki that writes itself as your agents work. One engine, not a patchwork.