​
State is not context

A transcript answers one question well: what was said next? An agent needs harder answers:
  • Which evidence supported this action?
  • Which file version is current?
  • Which preference superseded an older one?
  • What did this inference actually read?
  • What can leave the prompt without leaving history?
A chronological array can contain those facts, but it cannot express them directly. Meaning has to be recovered from position and prose.

​
The transcript tradeoff

A conventional loop stores:
To fit the next call, it usually chooses one of three policies: These are reasonable context policies. They are weak definitions of durable state.

​
The graph separation

StateWeave stores:
Parents mean direct dependency. resourceKey creates immutable version chains. The frontier marks current heads. A later turn begins from those heads rather than from the last line of a transcript. The next prompt is compiled from that graph. It includes required roots and heads, recent authoritative evidence, bounded operational ancestry, relevant older nodes, and a deterministic digest of the whole graph.
Separation

Persistence can remain lossless while model context remains bounded.

​
Why exact read-set parents matter

Suppose an agent has 200 graph nodes, but an inference receives 18. The runtime records those 18 node IDs as the inference’s parents. That gives causality a testable meaning: not “these nodes seem related,” but “these are the nodes the runtime rendered into this model call.” No second classification call is needed. No model-generated edge syntax is trusted. The lineage falls out of prompt construction.

​
Supersession without deletion

A file, preference, or durable lesson can change. StateWeave appends the new value and links it to the previous value through a stable resource key.
Projection treats the newest head as authoritative. Both versions remain in AgentState.

​
The boundaries of the claim

StateWeave does not make a model infallible. A graph can preserve bad evidence. Relevance ranking can omit useful history. A provider still processes the compiled prompt. Persistence and access control remain application responsibilities. The primitive improves state identity, provenance, boundedness, and inspectability. Product quality still depends on the model, tools, policies, and evidence you connect to it.