PALM Workshop · NeurIPS 2026 · Poster
Forecasting Retrieval-Bound Memory Failures, and Characterizing Its Limits
Redis
We embed conversation turns, memory texts, and queries with a shared analysis model. We measure query proximity, memory–conversation correspondence, and conversation structure.
We test 25 pre-answer features with a classifier trained on correctness labels. We evaluate four dense or hybrid retrieval implementations and one observation-summary comparator.
In our offline LongMemEval-small evaluation, geometry reaches a modest 0.640 AUROC across 2,367 question–system pairs. The question-difficulty baseline reaches 0.655.
Its strongest category score is 0.841 for single-session-assistant questions. System identity alone reaches 0.826 on the same category.
After controlling for question difficulty and system identity, geometry adds no statistically established pooled improvement.
ΔAUROC +0.002
95% CI [−0.010, 0.014]
This study does not establish a reliable live failure detector or identify why individual answers fail.
An interactive companion
This kind of visual exploration inspired the paper’s research questions.
Follow conversations as paths, compare sampled memories with source turns, and inspect their original-space similarities. Switch between UMAP, t-SNE, PCA, and available 3D views.
The Atlas uses shared analysis embeddings, not the systems’ native retrieval vectors. Similarity links are not verified provenance or fact coverage.
Zhechev et al. The Geometry of Agent Memory. PALM, NeurIPS 2026.
@misc{zhechev2026geometry,
title = {The Geometry of Agent Memory: Forecasting
Retrieval-Bound Memory Failures, and Characterizing Its Limits},
author = {Iliya Zhechev and Radoslav Ralev and Chris Coleman and
Srijith Rajamohan and Jen Agarwal and Tyler Hutcherson and Simba Khadder},
year = {2026},
eprint = {2610.26433},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26433}
}