Giving Loops Memory
For loops that run over many steps (research agents, multi-turn assistants), summarize prior steps into a short running memory instead of replaying the full history every time — this keeps context small and cheap..
The prompt
Here is a running summary of research so far: [summary]. Here is the newest finding: [new info]. Update the summary to under 150 words, keeping only what's still relevant to answering: [original question].
What to replace
Swap these placeholders for your own details before running the prompt:
[summary]your own value[new info]your own value[original question]your own value
Pro tip: Re-summarize memory every 3-5 steps in long-running agents to avoid context bloat and drifting off-task.
How to use this prompt
- Copy the prompt using the button above.
- Replace [summary], [new info], [original question] with your own details — the more specific you are, the better the output.
- Paste it into Claude and run it.
- If the answer feels generic, add constraints: audience, length, tone, and what to avoid. That single change fixes most weak output.
Learn the technique
Giving Loops Memory
Module 2 — Building Agent Loops · Loop & Agentic Engineering
Related prompts
Critique-and-Revise Loops — ChatGPT
A loop is simply feeding a model's own output back to it (or to a second prompt) for review and improvement, repeated until a quality bar is met.
ClaudeCritique-and-Revise Loops — Claude
A loop is simply feeding a model's own output back to it (or to a second prompt) for review and improvement, repeated until a quality bar is met.
ChatGPT / Agent toolsThe ReAct Pattern (Reason + Act)
ReAct loops interleave reasoning ("what should I do next?") with actions (calling a tool, searching, running code), then feed the result back in before reasoning again.
Any modelPlanner–Worker–Reviewer Teams
The most robust agent architecture is also the simplest: a Planner decomposes the goal into tasks, Workers execute one task each with narrow context, and a Reviewer checks results against acceptance criteria before anything ships.
Any modelLoop Budgets and Failure Handling
Production loops need budgets (max iterations, max cost, max time) and explicit failure paths: what happens when the Reviewer fails a result three times? Options: escalate to a human, fall back to a simpler method, or return a partial result flagged as unverified.
Any modelProject: Draft–Critique–Revise Content Machine
Deliverable: a working 3-role loop (writer, critic, editor) that produces a publishable piece of content on any topic you give it. Steps: 1.