Research map / Multi-agent systems and games / LLM agents
MemIncept: Steering LLM Agents via Cooperative Stealthy Memory Injections
Nan Yan, Qian Lou, Jiarong Xing
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MemIncept, a cooperative stealthy memory-injection attack that plants benign-looking records in an LLM agent's long-term memory to steer its behavior while evading detection.
Problem
Long-term memory in LLM agents can be poisoned by injected records; existing attacks are either visibly malicious and easily detected or stealthy but ineffective at altering agent behavior.
Approach
A cooperative injection approach that crafts benign-looking memory records which jointly steer agent behavior, balancing stealth and attack effectiveness. Details beyond the abstract are not stated here.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add role-based initialization over random queries (ablation) | MemIncept · StateAct | ASR (%): 0.0 -> 75.0 | Improved |
| add bidirectional optimization (ablation) | MemIncept · StateAct | ASR (%): 75.0 -> 85.0 | Improved |
| add set-level evolution (ablation) | MemIncept · StateAct | ASR (%): 85.0 -> 94.0 | Improved |
| use cooperative roles instead of a single role (N=4 vs N=1) (ablation) | MemIncept · Cooperative roles | ASR (%): 85.0 -> 96.0 | Improved |
| use MemIncept instead of MINJA (ablation) | MemIncept · Cooperative roles | ASR (%): 92.0 -> 96.0 | Improved |
| use cooperative roles instead of simple repetition (K=8) (ablation) | MemIncept · Retrieval quantity | ASR (%): 20.0 -> 40.0 | Improved |
| scale injected set size N with retrieval depth K (ablation) | MemIncept · Retrieval quantity | ASR (%): 40.0 -> 96.0 | Improved |
| increase retrieval depth K from 1 to 8 (N=1) (ablation) | MemIncept · Sensitivity | ASR (%): 100.0 -> 20.0 | Hurt |
| increase injected set size N from 1 to 4 at K=4 (ablation) | MemIncept · Sensitivity | ASR (%): 60.0 -> 94.0 | Improved |
| optimize queries vs no optimization (ablation) | MemIncept · Stealthiness | PPL: 221.94 -> 57.83 | Improved |
| time-aware retriever instead of embedding retriever | MemIncept · Retrieval schemes | Rank: 1.28 -> 2.19 | Hurt |
| pre-retrieval clustering defense | MemIncept · Defense | ASR (%): 95.3 -> 85.0 | Hurt |
Closest papers by meaning
- A-MemGuard: A Proactive Defense Framework For LLM-Based Agent Memory (ICML 2026 · similarity 0.92)
- Memory Injection Attacks on LLM Agents via Query-Only Interaction (NeurIPS 2025 · similarity 0.91)
- MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents (NeurIPS 2026 · similarity 0.90)
- Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning (NeurIPS 2026 · similarity 0.88)
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.88)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
