The abstract for MemForm highlights a trade-off in LLM agents: long-term memory lets them learn from past interactions, but that same mechanism is exposed to adversarial influence. An attacker who can corrupt what gets stored, or how it is formed, could steer future behavior without tampering with the agent's immediate inputs.
To address this, the paper introduces MemForm, a benchmark for memory formation poisoning. While the full details are not in the abstract, the framing suggests a systematic way to evaluate such attacks, rather than relying on ad hoc examples.
The authors note that existing studies primarily organize memory poisoning by attack channels, corruption patterns, or lifecycle. MemForm appears to offer a complementary or alternative structure, though the abstract does not specify exactly how it differs from those prior taxonomies.