# FlexA-Mem preprint reports gains in agent tasks under tight memory budgets

_Published Tuesday, October 27, 2026 at 11:07 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers propose FlexA-Mem, a memory system for language-model agents that reports an 11.5% improvement over the second-best method on AppWorld challenge tasks at a 4k memory budget. In a preprint, they also report a 12.2% improvement over A-Mem with a summary on 147 unused tasks with predictions fixed in advance.

The system stores each history unit once, then builds a compact timeline with relevant original text for each call without another language-model call. On LongMemEval and LoCoMo, it matches a dedicated keyword retriever that writes no memory, limiting the evidence for an advantage on exact-fact retrieval.

## Sources

- [cs.CL updates on arXiv.org](https://arxiv.org/abs/2610.33274)

---
Canonical: https://techandbusiness.org/newswire/-zQCq3qQjXeokLgPA6stUM
Published: 2026-10-27T15:07:44.253Z
Story chronology: 2026-10-27T04:00:00.000Z
Retrieved: 2026-10-27T17:51:52.298Z
Publisher: Tech & Business (techandbusiness.org)