Beyond similarity: trustworthy memory search for personal AI agents

The researchers propose MemGate, a lightweight trusted memory search plugin for personal AI agents, with only 9M parameters and a size of 35.1MB. The paper points out that existing memory pipelines mainly rely on semantic similarity retrieval; semantically similar memories may still be unsuitable for the current context, bringing threats such as cross-domain leakage, sycophancy, tool-call drift and memory-induced jailbreak. Evaluations on A-Mem, Mem0, MemOS and OpenClaw, a real agent environment with persistent state and tool-calling capabilities, show that long-term memory is not merely a utility layer but a persistent control channel that can reshape an agent's task understanding and action execution. MemGate is inserted between the vector memory store and the backbone LLM, using query-conditioned neural gating to admit candidate memories, without modifying the LLM, rewriting the memory store or using inference-time LLM review; across multiple memory frameworks, real agent scenarios and different LLM backbones, MemGate reduces memory-introduced threats while preserving long-term memory utility.