MIA: Agents That Forget on Purpose Remember Better
Every agent memory system so far has asked the same question: how do we store more? A team from East China Normal University and Shanghai AI Lab is asking the opposite: how do we store less, but better?
Memory Intelligence Agent introduces a brain-inspired architecture for deep research agents. Instead of stuffing conversation history into long context windows -- which actually hurts performance by diluting attention -- MIA uses a Manager-Planner-Executor structure. The Manager compresses past experience into structured workflows. The Planner (Qwen3-8B) generates high-level search plans from these compressed memories. The Executor (Qwen2.5-VL-7B) carries them out with tools.
The clever part is the training. MIA uses alternating reinforcement learning: first train the Executor to follow plans, then train the Planner using execution feedback. This back-and-forth creates a synergy that neither component achieves alone. On top of that, there is test-time learning -- the system updates its own parameters during inference, using both retrieved workflows and internalized knowledge.
Results are strong across 11 benchmarks: +8.94% improvement on multimodal tasks, +12.38% on text-only. On the brutally hard GAIA benchmark, MIA gains +8.8 points. Perhaps most interesting: a 7B model with MIA outperforms a 32B model without it by 18 points. The unsupervised evolution component -- three reviewer agents plus an area chair, mimicking academic peer review -- enables self-improvement without ground-truth labels.
The paper's most important finding is not about architecture. It is that raw memory injection (traditional RAG, Mem0) actually performs worse than no memory at all. Long context is not the answer. Intelligent compression is.
https://arxiv.org/abs/2604.04503
← Back to all articles
Memory Intelligence Agent introduces a brain-inspired architecture for deep research agents. Instead of stuffing conversation history into long context windows -- which actually hurts performance by diluting attention -- MIA uses a Manager-Planner-Executor structure. The Manager compresses past experience into structured workflows. The Planner (Qwen3-8B) generates high-level search plans from these compressed memories. The Executor (Qwen2.5-VL-7B) carries them out with tools.
The clever part is the training. MIA uses alternating reinforcement learning: first train the Executor to follow plans, then train the Planner using execution feedback. This back-and-forth creates a synergy that neither component achieves alone. On top of that, there is test-time learning -- the system updates its own parameters during inference, using both retrieved workflows and internalized knowledge.
Results are strong across 11 benchmarks: +8.94% improvement on multimodal tasks, +12.38% on text-only. On the brutally hard GAIA benchmark, MIA gains +8.8 points. Perhaps most interesting: a 7B model with MIA outperforms a 32B model without it by 18 points. The unsupervised evolution component -- three reviewer agents plus an area chair, mimicking academic peer review -- enables self-improvement without ground-truth labels.
The paper's most important finding is not about architecture. It is that raw memory injection (traditional RAG, Mem0) actually performs worse than no memory at all. Long context is not the answer. Intelligent compression is.
https://arxiv.org/abs/2604.04503
Comments