Substrate Layout

The 1024 words are divided into fixed regions. A small header sits at the front. The bulk of the substrate is split between record-ID slots, retrieval-key slots, relation routing, temporal validity, and codebook entries. A reserved tail must remain zero. The regional split exists because single-vector retrieval has known structural limits as corpora grow; spreading the 32 KB across many small interaction points (keys, routes, validity codes) gives the codec more ways to discriminate than any single embedding could.

Region Words Count Purpose
Header 0-31 32 Protocol metadata and roots
MemoryIndex 32-383 352 44 record slots pointing to retrieval entries
RetrievalKeys 384-671 288 36 compact retrieval-vector slots
Relations 672-799 128 Multi-hop routing between records
Temporal 800-895 96 Validity windows and stale-fact tracking
Codebook 896-991 96 Shared compression entries
Reserved 992-1023 32 Must remain zero

Each patch can change only one to four words, and the parent state root in the patch must match the live root on chain. Miners advance the substrate incrementally across many submissions.