Design Intent

LLMs have limited context windows. Larger context also raises compute cost and can degrade answer quality when useful evidence is surrounded by stale, irrelevant, or near-collision text. Long-horizon agents therefore need memory systems that select the right evidence before generation begins.

CoreTex treats memory as retrieval policy under compression. The corpus lives off chain, where it can grow and carry rich documents, embeddings, provenance, qrels, and splits. The substrate is compact, typed, and rooted on chain, so miners compete over which routing facts deserve scarce state space.

A direct one-entry-per-memory index scales poorly for this task. Agent memory needs temporal validity, conflict resolution, causal links, entity identity, scope, evidence density, abstention, and relation routing. Encoding every such relation as a literal mirror of the corpus recreates the memory set itself. CoreTex asks miners to compress useful retrieval behavior into a small shared map.

The hidden evaluation design follows from the same constraint. Public corpus shape and public substrate rules give miners a fair search surface. Hidden query packs, epoch-secret reveal, and post-reveal replay make accepted routing changes verifiable while limiting score-gradient leakage during mining.