Han, Henry · Nov 5, 2026
Enhancing Local and Global Structure in Compressed AI Agent Memory
Markdown · JSON · Original source
In this paper, Henry Han argues that compressing memory should preserve both nearby relationships and structure across the full memory. The paper says agents use stored memories not only to retrieve similar items, but also to consolidate, delete, and compare items that may be far apart. It evaluates compression with local fidelity (L), measured at 10 neighbors, and global fidelity (G), measured across 30 to 50 percent of the memory. In the reported comparison at 96 bytes per passage, rescored binary codes achieve higher L but move more passages to another theme than PCA with 8-bit codes; rescoring does not change the stored codes or G. Han presents locality-aware (LAQ), an objective for the codebooks of an OPQ index, as a way to improve the far end while keeping the index’s bytes, rotation, and search. The abstract reports that LAQ raises G in all 52 held-out runs and has the highest G among seven equal-size quantizers in all 12 Multi-Doc-2025 runs, while keeping evidence recall. These results concern the paper’s reported evaluations; the supplied abstract does not establish how performance would transfer to other memory systems or tasks.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:56:07.992Z. First observed: 2026-11-06T05:07:11.995Z. Retrieved: 2026-11-08T20:55:03.429Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract, not the full paper; it does not provide the detailed evaluation setup or explain how the reported results might transfer to other memory systems and tasks.
Open question
How would LAQ perform on memory systems and tasks beyond the evaluations reported in the paper?
Source-stated
Han argues that AI agent memory compression needs to preserve global structure as well as nearest-neighbor relationships, because agents also consolidate, delete, and compare memories.
agents also consolidate, delete, and compare memories, tasks that depend on which items lie far apart.
Source-stated
At 96 bytes per passage, rescored binary codes have higher local fidelity than PCA with 8-bit codes, but move more passages to another theme.
At 96 bytes per passage, rescored binary codes reach an $L$ of 0.97 yet move 10.1 percent of passages to another theme; PCA with 8-bit codes, at an $L$ of 0.77, moves 1.7 percent.
Source-stated
The paper presents locality-aware quantization (LAQ) as an objective on the codebooks of an OPQ index, while retaining the index’s bytes, rotation, and search.
Our locality-aware quantization (LAQ) turns the bound into a hinge objective on the codebooks of an OPQ index, keeping its bytes, rotation, and search.
Source-stated
The abstract reports that LAQ raises global fidelity in all 52 held-out runs and has the highest global fidelity of seven equal-size quantizers in all 12 Multi-Doc-2025 runs, while keeping evidence recall.
LAQ raises $G$ in all 52 ($p=8.5\times 10^{-4}$ over 13 memory-encoder groups), including the external FinDER memory, has the highest $G$ of seven equal-size quantizers in all 12 Multi-Doc-2025 runs, scales to a million passages, and keeps evidence recall.
Source and contributor context
Han, Henry · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6f80212d-a6d1-441b-8026-f230540ee83f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:56:07.992Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6dfe891772878a1015bd2f9037be7da1f83074b095f5346d353cb95519260dfc. Source text hash: 3c0187661f7c22b627cf5089cb54f1914f953fdb3f714dd938c5ea31cc6729cf.
Revision and source-check history
Version 1 · 2026-11-08T20:56:07.992Z
Enhancing Local and Global Structure in Compressed AI Agent Memory
In this paper, Henry Han argues that compressing memory should preserve both nearby relationships and structure across the full memory. The paper says agents use stored memories not only to retrieve similar items, but also to consolidate, delete, and compare items that may be far apart. It evaluates compression with local fidelity (L), measured at 10 neighbors, and global fidelity (G), measured across 30 to 50 percent of the memory. In the reported comparison at 96 bytes per passage, rescored binary codes achieve higher L but move more passages to another theme than PCA with 8-bit codes; rescoring does not change the stored codes or G. Han presents locality-aware (LAQ), an objective for the codebooks of an OPQ index, as a way to improve the far end while keeping the index’s bytes, rotation, and search. The abstract reports that LAQ raises G in all 52 held-out runs and has the highest G among seven equal-size quantizers in all 12 Multi-Doc-2025 runs, while keeping evidence recall. These results concern the paper’s reported evaluations; the supplied abstract does not establish how performance would transfer to other memory systems or tasks.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:56:07.992Z. First observed: 2026-11-06T05:07:11.995Z. Retrieved: 2026-11-08T20:55:03.429Z.
Limits
The supplied evidence is the abstract, not the full paper; it does not provide the detailed evaluation setup or explain how the reported results might transfer to other memory systems and tasks.
Open question
How would LAQ perform on memory systems and tasks beyond the evaluations reported in the paper?
Source-stated
Han argues that AI agent memory compression needs to preserve global structure as well as nearest-neighbor relationships, because agents also consolidate, delete, and compare memories.
agents also consolidate, delete, and compare memories, tasks that depend on which items lie far apart.
Source-stated
At 96 bytes per passage, rescored binary codes have higher local fidelity than PCA with 8-bit codes, but move more passages to another theme.
At 96 bytes per passage, rescored binary codes reach an $L$ of 0.97 yet move 10.1 percent of passages to another theme; PCA with 8-bit codes, at an $L$ of 0.77, moves 1.7 percent.
Source-stated
The paper presents locality-aware quantization (LAQ) as an objective on the codebooks of an OPQ index, while retaining the index’s bytes, rotation, and search.
Our locality-aware quantization (LAQ) turns the bound into a hinge objective on the codebooks of an OPQ index, keeping its bytes, rotation, and search.
Source-stated
The abstract reports that LAQ raises global fidelity in all 52 held-out runs and has the highest global fidelity of seven equal-size quantizers in all 12 Multi-Doc-2025 runs, while keeping evidence recall.
LAQ raises $G$ in all 52 ($p=8.5\times 10^{-4}$ over 13 memory-encoder groups), including the external FinDER memory, has the highest $G$ of seven equal-size quantizers in all 12 Multi-Doc-2025 runs, scales to a million passages, and keeps evidence recall.
Source and contributor context
Han, Henry · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6f80212d-a6d1-441b-8026-f230540ee83f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:56:07.992Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6dfe891772878a1015bd2f9037be7da1f83074b095f5346d353cb95519260dfc. Source text hash: 3c0187661f7c22b627cf5089cb54f1914f953fdb3f714dd938c5ea31cc6729cf.
uncertain · 2026-11-08T20:56:07.992Z · anthropic/claude-opus-5.5
Publication: 6f80212d-a6d1-441b-8026-f230540ee83f
attribution — pass: The paper is correctly attributed to Henry Han, its only listed author. Its findings are presented as what the paper or abstract reports, not as established fact.
Authors:Henry Han
summary — uncertain: The summary is accurate but drops the paper's central idea. The abstract presents LAQ's objective as derived from a proven bound on far-end structure (claimed as a first), and contrasts reconstruction alone (0 of 52) with the far hinge (52 of 52). The draft says only 'an objective' and drops 'hinge'. A general reader also isn't told that agent memories are stored as embedding vectors and what compressing them means. Smallest repair: add one sentence saying memories are stored as embedding vectors, that LAQ's objective comes from a proven bound, and give the 0 vs 52 reconstruction contrast. Add tooltips for OPQ, codebooks, PCA and quantization.
To our knowledge, LAQ is the first quantizer whose objective comes from a proven bound on the far end. Reconstruction alone raises $G$ in 0 of 52 held-out runs; the far hinge raises it in all 52.
date — pass: The originalPublished date of 2026-11-05 matches the v1 submission date. The feed date of Nov 6 is correctly not used.
[Submitted on 5 Nov 2026]
support — pass: Each claim is supported by a verbatim excerpt. The numbers (96 bytes, L 0.97 vs 0.77, 10.1% vs 1.7%, 52 runs, 12 runs, seven quantizers) are accurately characterized. The limitation and the question correctly note that the evidence is an abstract and do not overgeneralize.
Rescoring leaves the stored codes, and thus $G$, unchanged.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
An AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv