StateWeave: Choosing Unary, Pairwise, or Long-Context Agent Memory Under a Read Budget

Pairwise consistency and evidence access solve different failures in schema-defined agent memory. A violated dependency among linked beliefs can be repaired without another read; absent evidence cannot. StateWeave makes that distinction operational with provenance-linked unary marginals, bounded belief propagation activated by coupling, entropy spread, or verifier disagreement, and retrieval that begins from the implicated evidence. Across ConstraintQA and ToolAdapt, the adaptive mode improves Success over calibrated unary state by $2.3$ and $1.2$ points while reducing the cross-domain retrieval-trigger mean from $44.1$ to $37.3$. Equalizing supervision retains gains of $+0.9$ points (95% CI $[0.2,1.7]$) and $+0.6$ ($[0.1,1.3]$); MEMTRACK and StructMemEval transformations independently retain $+1.1$ and $+0.9$ same-supervision effects. The mechanism localizes sharply: on the pre-specified 29.4% ConstraintQA dependency slice, pairwise propagation adds $4.5$ points over calibrated unary state, compared with $0.6$ elsewhere. Against stronger alternatives, StateWeave-Adaptive leads published A-MEM by $1.2/0.8$ points and GraphMem-SQL by $1.6/1.0$ on ConstraintQA/ToolAdapt. It also exceeds LongCtx-128k by $2.3/1.7$ points using $7.6/6.9\times$ fewer tokens, whereas LongCtx-64k wins the diffuse-recall LoCoMo-Stream regime by $2.3$ points at roughly $4\times$ the token cost. A validation-frozen four-target diagnostic identifies both pairwise-positive domains and three of four candidate-set winners, making coupling and evidence spread measurable signals for allocating inference, retrieval, and context.