The Geometry of Consolidation

Summary

NeurIPS 2026 submission (third paper in a meaning-memory trilogy after The Price of Meaning and The Geometry of Forgetting). When a store replaces n cluster members with m<n representatives, identity error is lower-bounded by 1 - c1 m (θ'/d̄)^(d_eff/2) whenever the retrieval-cap half-angle θ' is smaller than mean within-cluster cosine distance d̄. GAC routes tight clusters (d̄ < θ') to the centroid and spread clusters to a residual-budgeted medoid; README claims Pareto dominance vs centroid/medoid/importance/selective-prune/PQ/OPQ/LSH/HNSW-prune across 6 encoders × 7 corpora × 10K→1M. Repo has paper/arxiv (48 pp) and paper/neurips, pip-installable gac/, experiments E1–E9, and results parquet. MIT. No public arXiv abs id yet (README cites arXiv submit/7412286 and submit/7411865 for the sister papers).

Keywords

gac · consolidation · semantic-memory · embeddings · neurips-2026 · sentra · geometry

Topics

semantic memory, embedding geometry, vector compression

Research notes

• Primary: GitHub README + API (MIT, Python, 110 stars / 9 forks at check). Authors Vangara (Sentra; Waterloo) and Gopinath (Sentra; MIT; corresponding). Discord posted paper/arxiv/main.pdf; cataloged the repo URL as the leftover GitHub item. No public arXiv abs for this paper at check. results/ parquet and public-corpus build scripts are experiment artifacts, not a new hosted pretraining corpus, so no datasets_local row.