← Back to the wire

How Much Memory Does Your Agent Actually Need?

AchievementResearchAug 18, 2026

A study across eight models found that agentic memory benefits depend on model capability, with strong models gaining from full guideline sets and weaker models performing best with selective retrieval. DeepSeek-V3.2 improved by 9.5 percentage points in task completion with full guidelines, while gpt-oss-120b gained 16.1 points using curated retrieval at only 5% additional token cost. The ALTK-Evolve framework distills reusable guidelines from prior trajectories without weight updates, and already-saturated models showed no measurable improvement.

Receipt № 14491 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

gpt-oss-120bModelDeepSeek-V3.2Model
Canonical: https://huggingface.co/blog/ibm-research/altk-evolve-hmm