# IBM and Hugging Face report model-specific gains from agent memory

_Published Tuesday, August 18, 2026 at 7:07 PM EDT · AI, Science · Latest · Tier 2 — Notable_

![IBM and Hugging Face report model-specific gains from agent memory — Primary](https://cdn-uploads.huggingface.co/production/uploads/6435a1131860001f144239ea/j-n1Au9SJ2-RGkT9i98u4.jpeg)

Hugging Face and IBM Research reported that ALTK-Evolve, which extracts and reuses guidelines from an agent's prior trajectories without weight updates, produced different results across eight tested models on the AppWorld benchmark. The report says curated retrieval raised gpt-oss-120b task completion by 16.1 percentage points with 5% more tokens, while full guideline sets worked better for some stronger models. Results are limited to AppWorld, and the authors say broader benchmarks and real-world deployments are in progress.

## Sources

- [Hugging Face](https://huggingface.co/blog/ibm-research/altk-evolve-hmm)

---
Canonical: https://techandbusiness.org/newswire/EphSCQ9Xibj-j-DZXl2AOC
Published: 2026-08-18T23:07:36.672Z
Story chronology: 2026-08-18T18:09:38.000Z
Retrieved: 2026-10-03T02:45:37.142Z
Publisher: Tech & Business (techandbusiness.org)