CoreTex

CoreTex is an evolving stack for agentic memory retrieval, purpose built from the ground up to integrate with existing BOTCOIN miner flow.

The stack consists of the corpus (the memory content itself), the substrate, (think of this as an expansive temporal map and routing layer), with state roots pinned onchain, the bespoke CoreTex retrieval router (decodes and interprets the substrate to surface relevant information), and Qwen 0.6B reranker model (scores against hidden/unseen memory eval tasks).

LLMs operate within limited context windows. As context fills, compute cost rises and answer quality severely degrades.

The agentic memory bottleneck is not the physical storage of information, but rather, how the LLM efficiently retrieves exactly what it needs, without stuffing the context window with a broad search.

Instead of asking miners to store more memories, CoreTex asks them to improve how memories are found. When a miner submits a substrate improvement, the CoreTex stack tests that update on unseen memory-search tasks. The small Qwen model evaluates and scores based on overall accuracy of context surfaced against current baseline. Changes that result in a score above a dynamically scaling threshold advance the onchain state, whereas weak, bogus, or stale updates are rejected.

The corpus evolves dynamically over time, just as real memory would. New information is injected, old information is rendered stale, conflicting information enters, various old hidden tests are retired, and new ones are added. Not only does this result in a more robust, generalized substrate composition over time, it acts as a lever to extend longevity by exposing new surface area for miners.

All context needed for a miner to submit valid substrate patches is accessible via the CoreTex skill file and miner facing endpoints -- however, anyone, whether mining or not, can also run an independent CoreTex validator client. This client is a direct copy of what runs on the coordinator server. Any state advancement can be replayed and validated by any independent party.

Although not required, running the CoreTex client locally allows a miner to screen/probe patches more rapidly, thus allowing them to submit only the most promising ones for onchain state advancements (more work, more reward).

The client itself is meant to be lightweight, and trivial to setup. Special care went into designing CoreTex to run on CPU rather than GPU, to ensure access was not walled behind expensive hardware. After installation, a one time setup command syncs the client with current state and validates any new state advancement (and optionally allows replay of any old state advancement).