Source
How Much Memory Does Your Agent Actually Need?
IBM Research evaluated how agentic memory, delivered as distilled behavioral guidelines, affects performance across eight different AI models. The study found that the optimal amount of memory depends on the specific capability level of the model.
Why it matters
Developers can increase AI agent accuracy and reduce operational costs by tailoring the volume of guidance to a model's strength rather than using a one-size-fits-all approach.
The details
- ALTK-Evolve distills reusable guidelines from prior trajectories without updating any model weights.
- gpt-oss-120b gained +16.1pp task completion using curated retrieval instead of full sets.
- DeepSeek-V3.2 saw a +9.5 percentage point increase in task completion with full guidelines.
Show entities and relationshipsHide entities and relationships
In this article
Key connections
GPT-5.5 uses ALTK-Evolve
GPT-5.5 was evaluated with ALTK-Evolve full guideline injection, gaining +2.9pp TGC.
DeepSeek-V3.2 competes with GLM-5
DeepSeek-V3.2 and GLM-5 are large mixture-of-experts models compared in the evaluation.
DeepSeek-V3.2 is built with Mixture of Experts
DeepSeek-V3.2 is a 671B Mixture of Experts model.
ALTK-Evolve competes with ACE
ALTK-Evolve is compared with ACE for delivering self-distilled agent guidelines.
ALTK-Evolve uses Agentic Memory
ALTK-Evolve implements agentic memory by distilling reusable behavioral guidelines from past trajectories.
ALTK-Evolve is related to AI Agents
ALTK-Evolve lets an agent learn from its own past trajectories by distilling reusable guidelines and injecting them back at inference time.
Show 11 more connectionsShow fewer connections
ALTK-Evolve uses AppWorld
ALTK-Evolve was evaluated across models using the AppWorld benchmark.
ALTK-Evolve uses Prompt Caching
ALTK-Evolve leverages prompt caching to keep static guideline sets affordable in production.
ALTK-Evolve uses ReAct Pattern
ALTK-Evolve injects self-mined guidelines at each ReAct Pattern step during inference.
gpt-oss-120b uses ALTK-Evolve
gpt-oss-120b was evaluated using ALTK-Evolve curated retrieval, achieving a +16.1pp TGC improvement.
DeepSeek-V3.2 uses ALTK-Evolve
DeepSeek-V3.2 was evaluated using ALTK-Evolve full guideline set, achieving a +9.5pp TGC improvement.
Claude Opus 4.6 uses ALTK-Evolve
Claude Opus 4.6 was evaluated with ALTK-Evolve full guideline injection, gaining +4.1pp TGC.
GLM-5 uses ALTK-Evolve
GLM-5 was evaluated with ALTK-Evolve, exhibiting saturated performance with 0.0pp gain.
Claude Opus 4.6 competes with GPT-5.5
Claude Opus 4.6 and GPT-5.5 are frontier proprietary language models compared in the evaluation.
GLM-5 is built with Mixture of Experts
GLM-5 is a 745B Mixture of Experts model.
gpt-oss-120b is built with Mixture of Experts
gpt-oss-120b is a 117B Mixture of Experts model.
DeepSeek-V3.2 uses ReAct Pattern
DeepSeek-V3.2 executes approximately 18 to 19 ReAct Pattern steps on average per task.
Related events
Study Demonstrates Agentic Memory Must Be Calibrated to Model Capability Tiers
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.