Approaches for Managing Agent Memory

Humans refine their skills and learn preferences through experience. But many AI agents lack this capacity for continual learning. Here, we give an overview of memory in the DeepAgents CLI. Here, we talk about two methods for updating memory: "explicit" updating from user-provided instructions and "implicit" updating from reflection over past DeepAgent sessions.

LangSmith fetch:
https://github.com/langchain-ai/langsmith-fetch

Claude Diary, an example reflection pattern:
https://rlancemartin.github.io/2025/12/01/claude_diary/

Deepagents-CLI:
https://github.com/langchain-ai/deepagents/tree/master/libs/deepagents-cli

Chapters:
0:00 - Introduction: Explicit vs. implicit memory updating
0:24 - Claude Code memory: The pound shortcut removal and natural language editing
0:54 - Deep Agents memory structure: Global and project-specific Agent.md files
1:31 - Demo: Explicit memory updating with Deep Agents CLI
2:24 - Introduction to implicit memory updating
2:50 - Claude Diary: A session-based memory system for Claude Code
3:18 - Research foundations: Generative Agents and Context Evolution papers
4:04 - How Claude Diary works: Diary and reflect commands
5:05 - Key insight: Memory refinement vs. append-only approaches
5:30 - Deep Agents session logging with LangSmith
6:24 - Understanding threads vs. traces in LangSmith
7:24 - Identifying implicit preferences in session history
7:54 - Demo: Using LangSmith Fetch to pull recent threads
9:06 - Demo: Reflecting on threads to extract and save preferences
10:44 - Results: Automatic memory update from implicit preferences
11:00 - Recap: Explicit and implicit memory patterns for agents