Hillock 0.6 Uses Bipolar Hypervectors to Retrieve Agent Memory, Skipping LLM Calls Below a Threshold

The open-source Hillock 0.6 project attempts to handle document extraction and first-stage retrieval without generative models. Its TALON pipeline uses Fastcoref, MiniLM, and GLiREL to convert text into subject–predicate–object triples, storing hard facts in SQLite. The strength of concept co-occurrences is maintained through gradient-free Hebbian updates, while conversations and queries are mapped into a 10,000-dimensional hypervector space whose elements are either +1 or −1. Only when the similarity of candidate facts exceeds a fixed threshold does the system pass the content to a local Ollama model to generate an answer; otherwise, it immediately returns a hard-coded refusal.

In version 0.6, HYDRA no longer compresses an entire query into a single vector. Instead, following ColBERT, it applies MaxSim between query tokens and fact tokens. To reduce the cost of comparisons in 10,000 dimensions, it first uses a 2,000-dimensional projection to eliminate most candidates before performing the full computation. Another feature, HYPERGRAPH-HDC, uses cyclic shifts to break the commutativity of Hadamard binding, encoding the order of two-hop and three-hop relationships into hypervectors. SQLite stores only single-hop triples, while multi-hop representations are stored as bit-packed BLOBs, preventing relationship paths from expanding directly into large numbers of rows.

The value of this design is that it makes “refuse to answer when evidence is lacking” part of the control flow rather than merely an instruction in the system prompt. Queries that cannot be answered never invoke the generative model. The developer says the complete pipeline uses less than 1.2 GB of VRAM on a GTX 1070, also supports CPU-only operation, and takes about 1.16 seconds in a quick benchmark of 32 queries.

The figures should not yet be generalized. The public test consists of just one 32-sentence document, 22 answerable questions, and 10 deliberately designed negative examples. Version 0.6 achieved extraction precision of only 13.8%, retrieval accuracy of 54.5%, overall gating accuracy of about 56.2%, and a negative-example blocking rate of 60%. A fixed threshold may therefore allow incorrect content through or reject valid questions. The next step should be to compare recall, false-rejection rates, and total computational cost across multi-domain corpora, different languages, and continuously updated memory, while also testing whether triple-extraction errors are amplified along multi-hop paths.