The same test, across a real HTTP boundary.
Memtru calls an authenticated API. A separate application owns the memory and a separate worker process handles delayed writes. The adapter has no direct database access.
What was measured
Captured locally on 23 October 2026: 20 of 20 intentionally vulnerable runs failed and 20 of 20 corrected runs passed. The timelines below are the last captured reports from that exercise, not a live scan or a customer deployment.
Two deterministic agent handlers keep separate logical memories in SQLite. The worker captures an update, waits while both copies are deleted, then resumes. The fixed worker checks a deletion tombstone inside its write transaction; the vulnerable worker ignores it.
Broken store
Marker: Not written yet- Press Next to start.
Fixed store
Marker: Not written yet- Press Next to start.
Additional checks in the exercise
- HTTP 503, timeout and a dropped connection produce ERROR, never PASS.
- A missing downstream positive control produces INCONCLUSIVE.
- Memory and held jobs survive application and worker process restarts.
- Scoped cleanup preserves unrelated synthetic records.
- The existing CLI returns exit 1 for the vulnerable version and 0 for the fixed version.
What this does not prove
The target is a synthetic application, without an LLM or A2A. Agents run as handlers inside the application process. This does not prove compatibility with every provider, production tenant isolation, physical erasure, power-loss recovery or every possible timing race.
Customer adoption requires an authorized sandbox, suitable retrieval paths and a supported lifecycle control. The hosted dashboard stores explicitly uploaded reports; it does not connect to arbitrary customer agents.
python -m pip install memtru==0.1.1
python prove.py --repeat 1 --output-dir ./reportsDownload the seven-file example and follow your first real agent test. No repository access or development dependencies are required. The example is synthetic; connecting your own sandbox needs the documented adapter mapping.