Griptape Tool: Prompt Summary

The Prompt Summary tool solves a problem every long-running agent eventually hits: context is a budget, and conversations blow through it. When an agent has been working for a while, its task memory is stuffed with full transcripts of earlier steps - every raw file read, every long intermediate answer. That's expensive (all those tokens) and it degrades quality (the model gets lost in the noise). This tool lets the agent compress its own memory: it summarizes what's in task memory, so the important stuff survives in a few hundred tokens instead of a few thousand.

Think of it as the agent's note-taking habit. Instead of re-reading the entire conversation to answer a question, the agent can summarize what it knows and work from the summary. In this pack's ecosystem that's the natural companion to long document-processing runs - the ones where you'd otherwise watch the context window fill up and start silently degrading the output. If you've ever had an agent forget the beginning of a task, this is the tool you were missing.

How it works

The node wraps Griptape's PromptSummaryTool with a PromptSummaryEngine. The engine does the summarizing, and the tool gives the agent a way to trigger it mid-task. Nothing runs at graph-build time - as with every tool in this pack, the agent decides when to reach for it.

The inputs that matter

This node is unusually lean. Where most tools in the pack have an off_prompt toggle, this one has none - the subclass removes it, so summarization runs in the agent's context by design (which makes sense: the agent needs to know its own summary). The one input is:

• prompt_driver - optional. The model that does the summarizing. This is the gotcha: leave it empty and the node builds the engine with no prompt driver, so you want to be sure a default prompt driver is available in your agent's setup. In practice, wire in the same prompt driver you're using for the agent (OpenAI, Ollama, whatever) so the summary is generated by a model you know is configured. If you're running everything local via Ollama, that works here too - summarizing is cheap work and doesn't need a flagship model.

One output, TOOL (TOOL_LIST), feeds an agent's tools.

Install

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

or ComfyUI Manager → search "Griptape" → install → restart.

Where people get burned

The main trap is expecting a summary to appear on the node. It won't - it's a tool the agent calls when it decides memory is getting heavy. If your agent never summarizes, it's usually a prompting problem: tell it explicitly to use the summary tool when the task is long, or when it needs to compress context before continuing.

The subtler trap: this tool is only useful when there's meaningful task memory to summarize. On a one-shot "answer this one question" workflow it does nothing, and you've added a tool for no reason. Reach for it on multi-step, document-heavy pipelines where context accumulates - that's where it earns its place. And if you're watching costs on a long agent run, a well-prompted summary tool is genuinely one of the best token-saving moves in this whole pack.