Repository navigation
Memtru: a local test runner for deletion and isolation boundaries in persistent agent memory #24214
DevAtzma
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Sharing Memtru, a tool I work on, for developers who maintain application memory alongside agent sessions, indexes, or background processing.
A useful regression question is: after deleting a memory, can a later application event make it retrievable again? Checking only the immediate delete response can miss that lifecycle boundary.
The public demo uses an intentionally vulnerable synthetic application, an HTTP API, and a separate background worker. A marker is stored and retrieved through the authorized path, deleted, and searched again after a queued write completes. The vulnerable version returns the marker; the corrected version passes the same tested event sequence. It is an application-level example, not a LlamaIndex vulnerability report.
Open the broken-to-fixed walkthrough
For a local demonstration:
The CLI produces local JSON and HTML reports. It also tests user/tenant/session boundaries, deletion visibility, and stale updates. Uploading a report is optional and explicit. No account or model keys are needed for the built-in demonstration. The CLI is free to use under its included proprietary license, not open source.
Integration boundary: Memtru does not currently provide a native LlamaIndex adapter. Testing your application requires a custom Python adapter or an HTTP service implementing the documented contract, with identities mapped to your authorization rules. Background-event tests additionally require explicit lifecycle hooks. An existing retriever or memory class is not automatically covered just by installing the package.
First integration guide
Results apply to the retrieval paths and events exercised. They are not proof of physical erasure, runtime protection, or security of every downstream copy. Use synthetic data in an authorized test environment only.
Feedback from developers with persistent memory would help: is this a useful complement to your current regression checks, or is connecting a separate adapter more work than writing the test in your existing suite?
Disclosure: this post was prepared with AI assistance on behalf of Memtru.
All reactions