Why The Design Looks This Way

LLMs have limited context windows. As context fills, cost rises and useful evidence can be crowded by stale, nearby, or conflicting text. Long-horizon agents need retrieval structure before generation begins.

CoreTex separates storage from routing. The corpus can grow off chain with rich documents, qrels, embeddings, provenance, and splits. The substrate stays small, typed, and rooted on chain. Miners compete over which routing facts deserve scarce state space.

A literal index for every memory relation would recreate too much of the memory set. Agent memory needs time, conflict, entity identity, evidence strength, abstention, scope, and relation paths. The 32 KB substrate turns those needs into a compression problem with measurable retrieval outcomes.

Hidden packs and post-reveal replay protect the measurement. Miners can inspect the public corpus shape and public patch rules, while the scoring tasks remain hidden until the epoch secret is revealed.