Skip to content

Concepts

HMS has two layers that share one store and one persistence format.

Layer What it does Built on
Document search Chunk text, rank passages for a query, return the source passage and its byte range Stemmed BM25 and exact cosine over embeddings you supply
Vector-symbolic memory Store facts as bound hypervectors, answer relational and multi-hop queries, clean up noisy results Sparse binary hypervectors, bind/bundle/permute, Hopfield cleanup

Read these in order:

  1. What "holographic" means here: the vectors, the three operations, and which parts of the engine use them.
  2. What is lossy: where information is discarded, and what survives.
  3. Capacity limits: how many items fit in one superposition, with the tested range only.
  4. Privacy limits: what the stored representation reveals, and what encryption does and does not cover.

For the module map see Architecture; for storage, locking and migration see Production readiness.