HRR
A distributed representation method based on phase vectors. It deterministically generates text atoms through SHA-256 hashing and supports the algebraic operations bind, unbind and bundle, and is used to represent entities and relations.
Timeline (last 90 days)
The seven-layer architecture includes: permanent memory (Markdown), fact storage (SQLite+FTS5+HRR+embeddings), technical wiki (Markdown), procedural memory (skills and scripts), history (SQLite and text search), compression (context summarization), and preservation (backup and restore).
HRR uses SHA-256 hashing to deterministically generate text atoms and supports algebraic operations such as bind, unbind, and bundle, but hash-generated vectors cannot reflect semantic similarity.
Measured on 45 real candidates, the similarity between the HRR query vector and the stored fact vectors fluctuated between -0.023 and +0.015, with a span of 0.038 constituting noise; after rescaling, the term's contribution to each score was 0.489–0.508, i.e., a near-constant 0.147–0.152 in the final result.
HRR's bundle() averages angles, making unbind and bundle non-commutative; when recovering 1, 2, 3, 5, 9, and 17 components from a bundle, the recovered values were all approximately 0.000.
The HRR fix: store the original text vector instead of the bundle and mark it in the database as textv1 to identify and migrate old rows; for topic matching, the similarity between a paraphrased query and the stored text vector was +0.328, versus +0.009 when not matching.
All Facts (5)
The seven-layer architecture includes: permanent memory (Markdown), fact storage (SQLite+FTS5+HRR+embeddings), technical wiki (Markdown), procedural memory (skills and scripts), history (SQLite and text search), compression (context summarization), and preservation (backup and restore).
50%UnverifiedHRR uses SHA-256 hashing to deterministically generate text atoms and supports algebraic operations such as bind, unbind, and bundle, but hash-generated vectors cannot reflect semantic similarity.
50%UnverifiedMeasured on 45 real candidates, the similarity between the HRR query vector and the stored fact vectors fluctuated between -0.023 and +0.015, with a span of 0.038 constituting noise; after rescaling, the term's contribution to each score was 0.489–0.508, i.e., a near-constant 0.147–0.152 in the final result.
50%UnverifiedHRR's bundle() averages angles, making unbind and bundle non-commutative; when recovering 1, 2, 3, 5, 9, and 17 components from a bundle, the recovered values were all approximately 0.000.
50%UnverifiedThe HRR fix: store the original text vector instead of the bundle and mark it in the database as textv1 to identify and migrate old rows; for topic matching, the similarity between a paraphrased query and the stored text vector was +0.328, versus +0.009 when not matching.
50%