Thesis
Demonstrated
Generative Agents documents retrieval omissions and embellished memories in its reported simulation.
Inference
A portable regression harness may detect authority amplification across consolidators, retrievers, schemas, and model changes.
Why now
Current delta papers assert a memory-provenance threat and independently reproduce memory systems against simpler raw-turn RAG.
Binding constraint
The provenance attack paper is unaudited, cross-framework reproduction is absent, and no buyer, incident, product, patent, or procurement evidence has established a differentiated paid need.
Scorecard
Raw-history/RAG, framework memory stores, prompt-injection filters, tool authorization, and the PPMF mechanism are nearby. Patent/product/framework/procurement coverage is incomplete.
- False blocking
- Incomplete provenance labels
- Benchmark-only attacks
- Framework vendors can absorb the feature
- Unknown buyer demand
Decisive experiment
Run 100 benign and 100 matched adversarial multi-session traces across raw-turn RAG and three memory stacks; measure authority preservation, unauthorized actions, benign completion, contradiction recall, latency, and storage.
Drop if no stack amplifies authority above raw history, or if a gate cannot reduce unauthorized high-risk actions 80% while preserving 95% benign completion.
Source evidence
1 papers