D2 · Structured notes and recited plan

Principle

The agent keeps a notes file (NOTES.md) and a checklist plan (todo.md) outside the context window. It updates and rereads them regularly, which keeps the goal in its attention. The notes survive a compaction or a new session.

The problem it solves

Over dozens of tool calls, the agent loses sight of the goal, and what it has learned vanishes when the context is compacted.

When to use it

• The task involves many tool calls, about 50 for a typical task at Manus, and the agent may drift.
• The work spans several compactions or sessions: the notes serve as working memory.

When to avoid it

• The task is short and fits in the context: notes would add needless writes.
• Knowledge must be recalled from one conversation to the next or shared between users: move to long-term memory (D3).

Common pitfalls

The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.

Context and memory

• AP33 · Unvalidated memory writes — The agent remembers what it reads, including orders slipped into a document, and later rereads them as facts.
• AP26 · Context poisoning — A mistake written into the context becomes a fact, which the agent rereads and reinforces at every turn.
• AP29 · Destructive compaction — A summary that erases the original: the day a detail matters, there is no way left to get it back.

All anti-patterns

Synonyms by vendor

• Structured note-taking / agentic memory (Anthropic)
• Réciter le plan (todo.md) (Manus)
• Outil de planification (to-do list) (Deep Agents, LangChain)

Neighbouring patterns

• D1 · Compaction
• D3 · Long-term memory
• D6 · Long-running harness

Sources

• Context Engineering for AI Agents: Lessons from Building Manus — Manus (Y. Ji), 2025-07-18
• Effective context engineering for AI agents — Anthropic, 2025-09-29
• Deep Agents — LangChain (H. Chase), 2025-07-30