The problem
Old facts and inferred preferences can look equally certain in an AI’s context. The user sees the wrong answer, but not the assumption that caused it.
Context Guard
Check what your
AI thinks it knows.
A stale fact can derail a useful answer. Review a few facts and assumptions, then see what changes.
Fictional persona. Scripted answers. Nothing saved or sent.
Batch 1 of 2
Did I get this right?
Skip or move on anytime. Unanswered items stay unreviewed.
Review complete
A clearer working picture.
Same task. Different context.
“Plan my next Cedar working session.”
Unchecked context
Confirmed context used
Changes live in this tab and reset when you reload. Dropping an item removes it from this demo’s working context, not from an AI account, chat history, or the fictional “before” snapshot.
Product notes
Memory needs
a review step.
An agent can remember the words and still get the person wrong.
The decisions
Show three items at a time. Separate stated facts from guesses. Make correction and removal as easy as confirmation. Keep unanswered items unreviewed, and show the effect on the same task.
This prototype uses the Context Guard review pattern. It edits a local working state, with no model call or account-memory access.
What I’d measure
Correction rate, facts dropped, and whether the revised answer fits the task better. Ask users how confident they feel in the agent before and after review. These are proposed measures, not collected results.