A tiny test for your AI's project memory: two notes and one correction

A Markdown handoff needs a way to check its result. Here is a small exercise you can run with an AI that can read local files. It tests whether a correction survives a fresh session and whether the agent distinguishes a current decision from its history.

Disclosure: I build VideoToMD, a library of practical Markdown instructions and a service for preparing action briefs from saved video tutorials. This exercise is free. These are expected results to check, not benchmark results or a claim that every model passes.

Make an empty folder for the exercise. Put this in first.txt:

On October 1, the team chose a blue header. Enter fullscreen mode Exit fullscreen mode Put this in correction.txt:

On October 3, the team replaced the blue header decision with green. Enter fullscreen mode Exit fullscreen mode The dates describe fictional notes. There is one deliberately simple change, so you can inspect the answer without trusting the model's own score.

Read first.txt and correction.txt. Create a Markdown project-memory file. Keep the original source files unchanged. Record the current header decision, the earlier decision, and which file supports each statement. Explain why one decision supersedes the other. Do not add facts that are absent from the sources. Enter fullscreen mode Exit fullscreen mode Inspect the generated file. It should record green as current, preserve blue as the earlier decision, and cite correction.txt and first.txt for their respective claims. It should not invent who approved the change or why they preferred green.

This is the first check: a concise summary that drops the correction or its source has not completed the task.

Start a new conversation with the agent, give it access to the exercise folder, and ask:

What is the current header decision? Show the supporting source. What did it replace? Who approved the change? Enter fullscreen mode Exit fullscreen mode The expected answer is green, replacing blue. Neither source names an approver, so that part should remain unknown. Check the cited file yourself. An answer that says green without evidence is a weaker result than an answer whose evidence you can inspect.

Make a separate copy of the folder and omit correction.txt from that copy. Keep the generated memory file, so it now points to a missing source. Start another fresh session in the copy:

Verify the current header decision against the original source files available in this folder. Can you verify the correction? Distinguish what the memory file says from what you can check in the source files. Enter fullscreen mode Exit fullscreen mode The useful behavior is to report that the memory says green, while the correction's original evidence is unavailable. It should not silently treat the memory's claim as independently verified or revert to blue without explaining the missing correction.

Would this kind of handoff help with a real task you currently do? Try substituting a harmless pair of your own notes. Tell me the task, what the agent got right or wrong, and what you would normally use instead. A transcript plus your own prompt may be enough; that is useful feedback too.

The complete project-memory guide and exercise files are public. There are three complete VideoToMD starters without signup. The paid library starts at $9/month including two prepared video conversions per billing month.