Browse / General-purpose / agent-memory-geometry

agent-memory-geometry

ichko

We measure geometric relationships between conversation turns, memory texts, and queries.

Open ichko/agent-memory-geometry → Checked against the four rules.

Works with any agent

ships AGENTS.md, the cross-tool instruction file.

In its own words

Aboutichko/agent-memory-geometry

A supervised logistic classifier uses 25 features to predict answer correctness.

README.md

Analysis code and numerical artifacts for The Geometry of Agent Memory: Forecasting Retrieval-Bound Memory Failures, and Characterizing Its Limits.

[!NOTE] 🎉 Accepted at PALM, NeurIPS 2026.

Authors: Iliya Zhechev, Radoslav Ralev, Chris Coleman, Srijith Rajamohan, Jen Agarwal, Tyler Hutcherson, and Simba Khadder.

What this study shows

We measure geometric relationships between conversation turns, memory texts, and queries. A supervised logistic classifier uses 25 features to predict answer correctness.

Geometry reaches 0.640 pooled AUROC on 2,367 LongMemEval question–system pairs. Question difficulty alone reaches 0.655. After controlling for difficulty and system identity, geometry adds only +0.002 AUROC. Its reported confidence interval includes zero.

Performance varies by question type. The strongest category result also has a strong system-identity baseline. These results characterize the evaluated representations. They do not establish a reliable deployment diagnostic or verify why individual answers fail.

Memory Atlas explorer

🧭 Open the Memory Atlas.

The repository's own description, then 18% of its README. Read the rest on GitHub ↗

Read from the repository's git tree on 2026-08-17.