Practical research tool

Context Thrash

Context thrash is repeated reading, searching, or reconsideration that consumes an AI agent’s context window without producing durable progress. It often appears as repeated tool calls, forgotten decisions, reopened files, and direction changes that force the agent to rebuild information it already had.

Try it now

Catch loops in an agent transcript

Paste a trace to surface repeated actions, reload language, and direction changes. Everything stays in your browser.

Live result

High thrash

67/100 thrash score across 9 events.

Repeated actions3
Reload signals9
Direction changes2
  • 3× read src/auth.ts
  • 2× read src/session.ts
  • 2× search for refresh token

What is context thrash in AI agents?

Built and reviewed by Imran
Reviewed 26 July 2026

How does it work?

  1. Paste a tool trace or transcript from an agent run.
  2. The analyzer counts repeated actions, reload language, and explicit reversals.
  3. Review the flagged loops and decide whether memory, planning, or tool feedback needs improvement.

When is it useful?

  • Debugging coding agents that repeatedly inspect the same files.
  • Finding wasted searches and forgotten decisions in long research runs.
  • Comparing memory or summarization strategies using the same task.

Example: the same file opened three times

A trace that reads auth.ts, searches for the login handler, and later repeats both actions is a useful thrash signal. The repetition may reflect weak memory, unclear tool output, or a plan that never records settled decisions.

What are the limitations?

  • Repeated work can be legitimate verification, so every flag needs human interpretation.
  • The analyzer uses transparent lexical signals rather than a learned semantic evaluator.

Questions about Context Thrash

Is context thrash the same as a long context window?

No. A long context window is capacity; thrash is inefficient reuse of that capacity. An agent can thrash in either a short or long window.

What causes context thrash?

Common causes include weak memory, vague plans, poor tool feedback, excessive handoffs, and summaries that omit decisions or constraints.

How can I reduce context thrash?

Persist decisions, make tool results easier to reference, summarize stable facts, and stop the run when repeated actions exceed a defined threshold.

One useful idea when the research moves. No noise.