Pinned post
Recent work on agent memory and “dreaming” 🧠💭 shows how memory helps agents improve through experience. But training this capability is challenging: we lack demonstrations of good memory behavior (what to remember, how to organize it, and when to retrieve it).
We study a self-improving recipe: the model discovers its own memory prompts, generates trajectories with the optimized prompts, and distills those behaviors into its weights.
🔗 facebookresearch.github.io/RAM/blogs/self…
facebookresearch.github.io Self-Improving Memory: Meta-Optimization and Harness Distillation for Continual Learning Agents Self-Improving Memory trains a model on its own discovered memory behaviors, so it uses the file system as memory under a generic harness without task-specific instructions. We study a self-improving recipe: the model discovers its own memory prompts, generates trajectories with the optimized prompts, and distills those behaviors into its weights.
🔗 facebookresearch.github.io/RAM/blogs/self…
Introducing: Self-Improving Memory 🧠✨📈View quoted post on X
We let the model discover its own memory skills, then train them into its weights:
- Meta-optimizes memory from the model's own runs with large gains.
- Distilling these gains works too.
Blog post: 📝 facebookresearch.github.io/RAM/blogs/self…
🧵1/6