Skip to content
Neural Daily

ResearchNews

FrontierMemory: can agents learn from their own past?

Letta's new benchmark does not test what an agent remembers but whether memory written from past runs helps solve new tasks. Opus 5.5 and GLM-5.3 lead at 73.3 percent.

By: Mikola Máté Patrik · · 2 min read

FrontierMemory by Letta wordmark over a grid of grey pixels
Image: Letta · Press kit · Official benchmark site social image (og:image) · Source

What changed?

Letta Research released FrontierMemory 1.0 (Letta). Older memory benchmarks such as LoCoMo and LongMemEval quiz agents on personal facts and are now largely saturated. FrontierMemory measures whether an agent can turn logs of past runs (trajectories) into memory that helps another agent solve new tasks.

How it is built

  • 15 tasks adapted from DeepSWE, SkillsBench, AutomationBench, APEX-Agents and EnterpriseOps-Gym so that solving them requires past experience.
  • Three kinds of memory: task-specific conventions, recurring user preferences, and non-obvious properties of the environment, such as which file holds the real limits.
  • Hundreds of prior trajectories are provided for learning, and only a minority are relevant.
  • During review, 53 percent of the kept source tasks had broken verifiers, which the team repaired.

The memory corpus is public on Hugging Face (frontier-memory-corpus).

Results

As memory writers, Opus 5.5 and GLM-5.3 tied for first at 73.3 percent, followed by Kimi K3 (66.7%), then GPT-6.1 Sol and Gemini 4 Argon (60%). GPT-6.1 Sol was by far the cheapest at $0.77 per task. Among harnesses, Letta's own Dream came out on top, which is worth reading with some caution since Letta also built the benchmark.

The most useful finding is practical: skill descriptions written as imperative "use when" instructions were actually loaded by the task agent, while generic descriptions were not. Overly specific rules, such as a condition keyed to a file size, failed to generalize to new tasks. If you build agents with memory or a skill library, both lessons apply right away.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles