Urgent.News

What's breaking now, across hundreds of outlets.

Tech

The Components of Agent Memory

The concept of memory in AI agents is a complex one that involves a division of labor, much like in the human brain. Human memory works by encoding experiences constantly, with the hippocampus capturing these moments for quick access. However, not all experiences make it to long-term memory, with consolidation happening during sleep and quiet rest. Agents, on the other hand, capture their work in logs or transcripts and can use idle time to reorganize their context, a process akin to "sleep-time compute."

This memory organization can be enhanced through a process similar to dreaming, where agents reflect on their logs to correct mistakes or create new notes and representations. To evaluate agent memory, researchers have tested systems like Claude, focusing on components such as dreaming, tagging salient events, and creating memory notes.

Claude's memory operates differently from the human brain, but it can be trained to recognize what's important by tagging salient events during a session. This tagging allows Claude to easily recall crucial information in the future. Moreover, Claude can act as a prediction system, writing down rules and theories about relevant patterns that it anticipates will be useful in future sessions.

This predictive capability has improved Claude's performance in tasks requiring long-term memory, such as Continual Learning Bench (CL-Bench) 1.0.

In this benchmark, Claude's ability to create learning loops by recording failures, testing theories, and forming rules has varied among different models. Claude Sonnet 5.5 and Claude Opus 5.5, for example, have been able to close the loop by capturing validated rules, answers to earlier questions, and warnings about potential traps.

This approach to memory has led to improved performance in tasks like playing Pokémon Red, where agents must navigate complex, long-horizon tasks and make judgments about what information to retain in memory.

Rewritten by urgent.news; may contain errors — check the original.

Other reports of this story

Every outlet that reported this, and when →

Read the original at anthropic.com →

Report a problem with this story

More in Tech

Treat a crawler skill's output as a versioned interface

A saved crawler workflow is easier to reuse when the next person can trust its output shape. Finding the same button again is only part of the job.

  • Define contract before saving skill to prevent downstream issues
  • Specify key fields: source ID, URL, email handling, row limits
  • Use Python validator to check contract compliance

Ming-3.0-tts on a 16 GB card: the GPU sat mostly idle, and why a bigger one barely helps

When a speech model runs slower than real time, the reflex is to rent a bigger GPU. With Ming-3.0-tts-0.5B that buys very little. An RTX 2000 Ada sat idle more than two thirds of the time.

  • RTX 2000 Ada GPU idle >2/3 time while processing Ming-3.0-tts
  • RTX 4090 only effective once in five instances, 17% speedup
  • 16 GB card insufficient; 4 GB estimated minimum for stable performance

My Analytics Agent Stopped Repeating Its Wrong Guesses With Hindsight

My Analytics Agent Stopped Repeating Its Wrong Guesses With Hindsight Every Monday morning, someone in leadership asks why a regional revenue metric dropped, and an analyst spends two hours checking…

  • WhyDesk replaces LLM-driven tool with Hindsight methodology
  • System uses Vectorize agent memory to prioritize efficient diagnostics
  • Three memory types: world, experience, observation

More from Saturday 7 November →