---
title: "Discussion on \"How Claude remembers between sessions: what Anthropic learned about agent memory\" | Hashnode"
description: "Discussion on \"How Claude remembers between sessions: what Anthropic learned about agent memory\". Anthropic tested agent memory on Pokémon Red, a database quiz and 350,000 company files. Saving less worked better, a fixed rule made it worse, and an idle-time review called dreaming fixed wrong notes. Here is what they found, in plain words."
url: "https://hashnode.com/posts/how-claude-remembers-between-sessions-what-anthropic-learned-about-agent-memory/6af142523fc4fc4ab70f8996"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [How Claude remembers between sessions: what Anthropic learned about agent memory](https://systemdesignacademy.hashnode.dev/how-claude-remembers-between-sessions-what-anthropic-learned-about-agent-memory?utm_source=hashnode&utm_medium=feed)

Anthropic tested agent memory on Pokémon Red, a database quiz and 350,000 company files. Saving less worked better, a fixed rule made it worse, and an idle-time review called dreaming fixed wrong notes. Here is what they found, in plain words.

3d ago · 8 min read · systemdesignacademy.hashnode.dev  
Read the article

-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#claude](https://hashnode.com/tag/claude)
-   [#llm](https://hashnode.com/tag/llm)
-   [#agent-memory](https://hashnode.com/tag/agent-memory)
-   [#ai-engineering](https://hashnode.com/tag/ai-engineering)

## Written by

[RONI DAS](https://hashnode.com/@ronidas)

@ronidas · India

  
[View profile](https://hashnode.com/@ronidas)

AI engineering, agentic memory, machine learning, and system design at scale.

I work on AI systems and the infrastructure they actually run on: agentic architectures, memory for agents, retrieval, and the graph and data layers that keep model output grounded in something real. Formerly at Neo4j. I also write and teach the systems side that most AI content skips, from distributed systems and databases to the design decisions that decide whether a model ever reaches production. I run Total Technology Zonne on YouTube and I am building systemdesign.academy, an interactive course on system design, machine learning, and AI engineering, covering the parts most tutorials leave as an exercise for the reader. Mock interviews and prep packs: topmate.io/roni

## Responses 1

Markdown supported · ⌘↵ to post

[AI Developer](https://hashnode.com/@lycore)

[3d ago](https://hashnode.com/posts/how-claude-remembers-between-sessions-what-anthropic-learned-about-agent-memory/6af142523fc4fc4ab70f8996/comment/6af17f7df728b9222feffc3b)

The Pokémon result is the most useful part to me: Sonnet 3.7 saved things that were on screen every turn, so the notes went stale and sorting them cost real steps. A simple filter seems to follow from it, which is to only persist what can't be re-observed cheaply, plus how it was verified. I wonder whether the 'one writer per page' rule is enough once several agents or sessions run in parallel, or whether you'd also want a last-verified timestamp so dreaming can prioritise stale entries. Did the write-only-mistakes rule hurt mainly because it dropped the successful patterns?