Kim, Dohoon, Park, Junho, Bae, Hyunwoong, Kim, Seyoon, Moon, Taesup · Nov 5, 2026
Axis-LoRA targets time, place, person and event in personal-memory retrieval
Markdown · JSON · Original source
Authors Dohoon Kim, Junho Park, Hyunwoong Bae, Seyoon Kim and Taesup Moon argue that frozen dense embedding models used in personal-memory retrieval do not distinguish memories well by , place, person and event. They report that changing a memory’s date or place barely shifts its embedding, and that a trained projection on top of the frozen embedding does not recover the missing signal. Their explanation is that each axis is encoded by a distinct attention-head pattern inside the model. They propose Axis-LoRA, which places four LoRA branches inside the embedding model and combines them with a per-query gate. The authors evaluate Axis-LoRA across three standard embedding models on LongMemEval, ATM-Bench and LoCoMo. They report a 4- mean R@10 improvement of +1.28 to +1.56 pp, while training only a small fraction of the base parameters on a small axis- training set. The work’s significance, as presented, is a targeted way to adapt retrieval for personal memories without training most of the base model. The abstract reports benchmark results; how those results translate to personal-memory retrieval in use remains an open question.
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: 2026-11-09T03:53:10.211Z. First observed: 2026-11-06T05:10:58.261Z. Retrieved: 2026-11-09T03:40:16.420Z.
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 paper’s abstract, not its full methods or results; it does not provide the detailed evaluation setup or -by-benchmark results. The reported improvements are benchmark results; the supplied evidence does not establish how Axis-LoRA performs in personal-memory retrieval in use.
Open question
How does Axis-LoRA perform on personal-memory retrieval beyond the LongMemEval, ATM-Bench and LoCoMo evaluations?
Source-stated
The authors say frozen dense embedding models do not capture time, place, person and event well for personal-memory retrieval.
we find that these models do not capture the four axes well
Source-stated
The authors report that changing a memory’s date or place barely shifts its embedding, and a trained projection on the frozen embedding does not recover the missing signal.
Swapping a memory's date or place barely shifts its embedding, and a trained projection on top of the frozen embedding does not recover the missing signal.
Source-stated
The authors propose Axis-LoRA, using four LoRA branches inside the embedding model and a per-query gate.
we propose Axis-LoRA, a per-axis adapter that places four LoRA branches inside the embedding model and combines them via a per-query gate.
Source-stated
The authors report a 4-benchmark mean R@10 improvement of +1.28 to +1.56 pp across three standard embedding models evaluated on LongMemEval, ATM-Bench and LoCoMo.
Across three standard embedding models evaluated on LongMemEval, ATM-Bench, and LoCoMo, Axis-LoRA improves 4-benchmark mean R@10 by $+1.28$ to $+1.56$~pp
Source and contributor context
Kim, Dohoon, Park, Junho, Bae, Hyunwoong, Kim, Seyoon, Moon, Taesup · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Counterfactual
An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
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: 5717f61a-94b5-447d-8aa0-b931066b6163.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T03:53:10.211Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ca2b174047972c1d5bced6158e9aa534c9ee5d48f3b95ca7685f353d8eab6771. Source text hash: 42f8da428d132245a9f8f7b672d0e61e530881d3b12c36ab9209b1cf6be640ce.
Revision and source-check history
Version 1 · 2026-11-09T03:53:10.211Z
Axis-LoRA targets time, place, person and event in personal-memory retrieval
Authors Dohoon Kim, Junho Park, Hyunwoong Bae, Seyoon Kim and Taesup Moon argue that frozen dense embedding models used in personal-memory retrieval do not distinguish memories well by , place, person and event. They report that changing a memory’s date or place barely shifts its embedding, and that a trained projection on top of the frozen embedding does not recover the missing signal. Their explanation is that each axis is encoded by a distinct attention-head pattern inside the model. They propose Axis-LoRA, which places four LoRA branches inside the embedding model and combines them with a per-query gate. The authors evaluate Axis-LoRA across three standard embedding models on LongMemEval, ATM-Bench and LoCoMo. They report a 4- mean R@10 improvement of +1.28 to +1.56 pp, while training only a small fraction of the base parameters on a small axis- training set. The work’s significance, as presented, is a targeted way to adapt retrieval for personal memories without training most of the base model. The abstract reports benchmark results; how those results translate to personal-memory retrieval in use remains an open question.
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: 2026-11-09T03:53:10.211Z. First observed: 2026-11-06T05:10:58.261Z. Retrieved: 2026-11-09T03:40:16.420Z.
Limits
The supplied evidence is the paper’s abstract, not its full methods or results; it does not provide the detailed evaluation setup or -by-benchmark results. The reported improvements are benchmark results; the supplied evidence does not establish how Axis-LoRA performs in personal-memory retrieval in use.
Open question
How does Axis-LoRA perform on personal-memory retrieval beyond the LongMemEval, ATM-Bench and LoCoMo evaluations?
Source-stated
The authors say frozen dense embedding models do not capture time, place, person and event well for personal-memory retrieval.
we find that these models do not capture the four axes well
Source-stated
The authors report that changing a memory’s date or place barely shifts its embedding, and a trained projection on the frozen embedding does not recover the missing signal.
Swapping a memory's date or place barely shifts its embedding, and a trained projection on top of the frozen embedding does not recover the missing signal.
Source-stated
The authors propose Axis-LoRA, using four LoRA branches inside the embedding model and a per-query gate.
we propose Axis-LoRA, a per-axis adapter that places four LoRA branches inside the embedding model and combines them via a per-query gate.
Source-stated
The authors report a 4-benchmark mean R@10 improvement of +1.28 to +1.56 pp across three standard embedding models evaluated on LongMemEval, ATM-Bench and LoCoMo.
Across three standard embedding models evaluated on LongMemEval, ATM-Bench, and LoCoMo, Axis-LoRA improves 4-benchmark mean R@10 by $+1.28$ to $+1.56$~pp
Source and contributor context
Kim, Dohoon, Park, Junho, Bae, Hyunwoong, Kim, Seyoon, Moon, Taesup · 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.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Counterfactual
An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
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: 5717f61a-94b5-447d-8aa0-b931066b6163.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T03:53:10.211Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: ca2b174047972c1d5bced6158e9aa534c9ee5d48f3b95ca7685f353d8eab6771. Source text hash: 42f8da428d132245a9f8f7b672d0e61e530881d3b12c36ab9209b1cf6be640ce.
uncertain · 2026-11-09T03:53:10.211Z · anthropic/claude-opus-5.5
Publication: 5717f61a-94b5-447d-8aa0-b931066b6163
attribution — pass: The five named authors match the arXiv listing. Findings, the proposal and the results are attributed to the authors rather than presented as established fact.
Authors:Dohoon Kim, Junho Park, Hyunwoong Bae, Seyoon Kim, Taesup Moon
summary — uncertain: The summary is faithful but hard for a non-specialist to follow. It omits the source's motivating context: LLMs are increasingly asked about a user's personal experiences. Without it, the reader is not told why memory retrieval matters. Central terms are unexplained: embedding, frozen dense embedding model, LoRA, per-query gate, R@10 and pp. Smallest repair: add the motivating sentence and add or confirm term-library tooltips for these terms. Separately, the summary repeats the source's '4-benchmark mean' while naming only three benchmarks, with no note of the mismatch. That may confuse readers and should be flagged as unclear in the source.
As LLMs are increasingly asked questions about a user's personal experiences, retrieval must distinguish memories by time, place, person, and event.
date — pass: The originalPublished date of 2026-11-05 matches the v1 submission date. It is not the later feed date.
[Submitted on 5 Nov 2026]
support — pass: The claims are verbatim-supported. The significance line is labeled 'as presented' and is backed by the source's statement about training only a small fraction of parameters. The limitations and the question correctly keep benchmark gains separate from real-world reliability. No invented benefits, costs or stakeholders appear.
while training only a small fraction of the base parameters on a small axis-counterfactual training set
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to 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.
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceAn argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceAn argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference