Anthropic proposes organising AI agents' memory into four layers with a periodic review
An Anthropic research text describes how to give an agent memory between sessions, and a teacher in the group has adapted their own system to that idea and explains it in an article with instructions.
AI agents for the computer, such as Claude Code, remember nothing from one session to the next unless they are given some way of doing so. On 6 November Anthropic published “The components of agent memory”, which arranges that memory into four layers: transcripts, a searchable record of what the agent has done; memory notes, which the agent writes during sessions; representations, ordered structures built from the notes, such as a wiki of linked pages; and injected context, a small layer loaded at the start of each session that must be kept short because overly long instructions reduce how well the agent follows them.
The text adds an offline review, which it calls “dreaming”: the agent goes over its transcripts to cross-check the notes, correct errors, build representations and decide what is preloaded in future sessions. The authors warn that this review consumes computing power and should be focused on what is likely to change.
Juan José de Haro, author of the blog Bilateria, has adapted his personal system to this structure and explains it in the article “Agentes de IA que aprenden mientras duermen” (AI agents that learn while they sleep), published on 7 November. He proposes a short profile, agent notes with an index, a project wiki with decision logs and the saved conversations as raw material, with a daily review that only writes a proposal: the agent checks each point before applying it and changes to the profile require the person’s approval. It includes a prompt for asking the agent itself to set the system up.
The article points out limits that matter to anyone weighing it up for a school: saved conversations contain sensitive information, the review can make mistakes, and each tool’s memory is not shared with the others. On what student data an agent should be able to see and remember, the paper has already published a report.
Anthropic proposes organising AI agents' memory into four layers with a periodic review
An Anthropic research text describes how to give an agent memory between sessions, and a teacher in the group has adapted their own system to that idea and explains it in an article with instructions.
AI agents for the computer, such as Claude Code, remember nothing from one session to the next unless they are given some way of doing so. On 6 November Anthropic published “The components of agent memory”, which arranges that memory into four layers: transcripts, a searchable record of what the agent has done; memory notes, which the agent writes during sessions; representations, ordered structures built from the notes, such as a wiki of linked pages; and injected context, a small layer loaded at the start of each session that must be kept short because overly long instructions reduce how well the agent follows them.
The text adds an offline review, which it calls “dreaming”: the agent goes over its transcripts to cross-check the notes, correct errors, build representations and decide what is preloaded in future sessions. The authors warn that this review consumes computing power and should be focused on what is likely to change.
Juan José de Haro, author of the blog Bilateria, has adapted his personal system to this structure and explains it in the article “Agentes de IA que aprenden mientras duermen” (AI agents that learn while they sleep), published on 7 November. He proposes a short profile, agent notes with an index, a project wiki with decision logs and the saved conversations as raw material, with a daily review that only writes a proposal: the agent checks each point before applying it and changes to the profile require the person’s approval. It includes a prompt for asking the agent itself to set the system up.
The article points out limits that matter to anyone weighing it up for a school: saved conversations contain sensitive information, the review can make mistakes, and each tool’s memory is not shared with the others. On what student data an agent should be able to see and remember, the paper has already published a report.
