How Much Memory Does Your Agent Actually Need?

THE NEWS

What happened

IBM Research has released ALTK-Evolve, a breakthrough in AI agent memory management. The study proves that 'more memory' isn't always better; it depends on the model's capability. Strong models benefit from full guideline sets, while weaker models achieve peak performance with curated retrieval of task-specific lessons. This method improves reliability and task completion without updating model weights or requiring human annotation. Crucially, prompt caching keeps production costs low. The findings were validated on AppWorld, a rigorous benchmark of multi-step tasks.

AT A GLANCE

Key facts

The main verified details from the original report are summarized below.

• ALTK-Evolve allows agents to learn from past trajectories by distilling reusable guidelines without weight updates or human annotation.
• Strong models with headroom (e.g., DeepSeek-V3.2) perform best when given the full guideline set, gaining up to +9.5 percentage points in task completion.
• Weaker models are 'drowned' by large guideline sets; they achieve peak performance (+16.1pp for gpt-oss-120b) using a compact core plus per-task curated retrieval.
• Saturated models show no measurable gain from additional memory, indicating a ceiling on current task capabilities.
• Curated retrieval offers the best balance of accuracy and cost, increasing token usage by only ~5% for significant performance gains.
• Prompt caching is critical for production efficiency, as static guideline sets can be cached to reduce inference costs substantially.
• The method was validated on AppWorld, a benchmark featuring 585 multi-step tasks across nine simulated applications.