Structured note-taking, also called agentic memory, is a technique in which an agent regularly writes notes that are persisted to memory outside of the context window and pulled back into it at later times. Anthropic presents it as providing persistent memory with minimal overhead: the pattern lets an agent track progress across complex tasks and hold on to critical context and dependencies that would otherwise be lost across dozens of tool calls.
Usage
Anthropic gives two shapes for the pattern: Claude Code creating a to-do list, and a custom agent maintaining a NOTES.md file. It also points to Claude playing Pokémon as a demonstration of the technique outside coding domains, where the agent maintains precise tallies across thousands of game steps — tracking, for example, how many steps it has spent training in one area and how many levels that has gained toward a target. Anthropic reports that without any prompting about memory structure the agent develops maps of the regions it has explored, remembers which key achievements it has unlocked, and maintains strategic notes on which attacks work best against different opponents. After context resets it reads its own notes and continues multi-hour training sequences or dungeon explorations, which Anthropic describes as enabling long-horizon strategies that would be impossible while keeping all the information in the context window alone.
Anthropic has also released a memory tool in public beta on the Claude Developer Platform, which stores and consults information outside the context window through a file-based system. It says this lets agents build up knowledge bases over time, maintain project state across sessions, and reference previous work without keeping everything in context.
A later Claude Cookbook notebook,
Context engineering: memory, compaction, and tool clearing, treats memory in this sense
as the primitive for just-in-time retrieval (see Just-in-time context retrieval):
rather than loading everything up front, the agent stores what it learns and pulls it back on demand.
It presents the memory tool as memory_20250818, a client-side tool — the model issues memory
operations (view, create, str_replace, insert, delete, rename) and the application executes them
against storage it implements and controls — and says the API auto-injects a system prompt telling the
model to check its memory directory before doing anything else and to assume its context may be reset at
any moment. In the notebook’s demonstration, a second session given the first session’s saved memory
files opened by reading them and built on them, while the same follow-up task with an empty memory
directory had to go back to the source documents to rediscover the same facts.
The notebook also relays several ways to shape what gets saved: a system-prompt instruction limiting memory to a given topic; asking the model to keep its memory folder up to date, coherent and organized rather than accumulating new files; running a dedicated first session that sets up memory artifacts such as a progress log and a feature checklist for later sessions to read; and client-side storage hygiene — capping file sizes, clearing out files not accessed for a long time, and guarding against path traversal.
When It Applies
- Applies to long-horizon work with clear milestones. Anthropic says note-taking excels for iterative development, in contrast to Compaction, which it recommends for tasks requiring extensive back-and-forth.
- Assumes a persistence location outside the context window and a mechanism for reading the notes back in — a file the agent writes and re-reads, or a platform memory tool.
- The Claude Cookbook notebook scopes the memory tool to the cross-session problem: it gives lossless fidelity on whatever the agent chose to save, but does nothing about context growth within a session and adds tool-call overhead for every read and write. It suggests skipping memory where every session should start fresh, such as a chatbot whose conversations should be independent, and flags stale memory when facts change, and a data policy for personal or sensitive information when the agent learns user preferences, as things to watch.
- Established as Anthropic’s own practice, evidenced by Claude Code’s to-do list, the Claude playing Pokémon demonstration, and a memory tool in public beta. The Pokémon account is a demonstration Anthropic reports rather than a controlled measurement.