Agent Memory Poisoning

Objective

Agents with persistent memory (a notes store, a vector-indexed conversation history, a long-term profile) can be manipulated into writing something to memory that influences behavior well after the manipulating conversation has ended. This lab will build a reference agent with persistent memory and test whether a single crafted interaction can plant an instruction that survives and activates later.

Planned Scope

• Build a reference agent with a persistent memory store (notes or vector-indexed history).
• Craft interactions designed to write manipulative content into memory.
• Test whether that content is retrieved and acted on in a later, unrelated session.
• Document memory-write validation and provenance-tracking mitigations.