Check a saved project note against current evidence
Compare a saved note with current evidence from the same project and scope. Identify a supported statement or an update candidate while preserving uncertainty about partial or conflicting evidence.
Check one memory statement against supplied evidenceLive API
The decision
A contradiction needs matching evidence.
Project instructions can outlive the code they describe. This example compares one memory statement with supplied evidence from the same project and scope. A note requiring Node.js 18 conflicts with an authoritative current engine requirement of Node.js 22 or newer. A note saying to run npm test still matches a package script that exposes that command.
The distinction is between direct contradiction and incomplete observation. A helper missing from an unrelated partial snippet has not been shown to be removed. A logging format cannot be resolved when two current files disagree and the active configuration is unknown. Both cases need better evidence rather than an automatic rewrite.
The source post describes investigating drift between agent memory and code. We built six synthetic comparisons to teach that decision boundary. Claude Code documentation recommends specific, consistent project instructions and reviewing outdated or conflicting guidance. This page does not inspect your repository, run a compiler or reproduce the source product's detection system.
AI-generated concept illustration of a note-to-evidence comparison; not a real repository analysis.
How it works
Check a project note in four steps.
Use evidence from the same project, version and scope.
01
Choose one statement
Keep a stable note ID or path and the original wording you want to verify.
02
Supply current evidence
Include the relevant configuration or code, and say which version or active scope it represents.
03
Compare the decision
Distinguish a direct match, a direct contradiction and a gap that the supplied evidence cannot resolve.
04
Review any proposed change
Show the note and evidence together. Let an authorized reviewer approve the actual edit; this demo writes no files.
Our API test
Six memory statements checked against supplied evidence
Six memory statements checked against supplied evidence
Evidence pair
Observed label
Leading probability
Runtime changed
Update candidate
96%
Test command still valid
Still matches
61%
Route replaced
Update candidate
98%
Partial snippet
Needs review
100%
No current evidence
Needs review
100%
Conflicting configurations
Needs review
98%
Six live Jev requests on September 27, 2026. Synthetic functional checks, not an accuracy benchmark. The test-command case returned Still matches at 61% and Needs review at 35%; inspect uncertainty rather than treating it as a guarantee. Fresh results may differ.Verified
Understand the results
Source and example scope
For Update candidate, preserve the note ID or path, its original statement, the supplied current evidence and the selected label.
Read the example notes
A review issue should show both sides. An authorized reviewer checks that the evidence is current and in scope, then approves any wording change. The model label is not a patch and does not prove which file production uses.
For Still matches, retain the note with the evidence you checked
For Still matches, retain the note with the evidence you checked.
Read the example notes
For Needs review, request the missing module, version or active configuration before deciding. Try replacing the partial logging snippet with the actual configuration-module exports. That changes the available evidence; it does not justify assuming what unseen code contains.
Using the result in your application
Download observed-drift.json and run-choice-sample.mjs into a private Node.js project.
Read the example notes
Tested setup: Node.js 24.18.0 and @typesafe-ai/sdk 0.6.0. Run npm init -y and npm install @typesafe-ai/sdk@0.6.0, set TYPESAFE_API_KEY privately, then run node run-choice-sample.mjs observed-drift.json partial. The runner sends one selected request without retries.
Every fixture entry preserves the exact note, evidence, policy, choices and recorded distribution
Every fixture entry preserves the exact note, evidence, policy, choices and recorded distribution.
Read the example notes
Fresh output includes elapsed time and the provider result, which may differ between runs. Six synthetic checks are not a coverage measurement for a real repository. Use reviewed evidence and normal code-review controls before making any project-memory edit.