A learned explicit-memory agent
Explicit Memory is our second research paper on AI systems with explicit memory. This release takes the original Human-Like Memory Systems direction one step further: instead of hand-writing the memory policy, we train the agent to decide what should be forgotten, stored episodically, or stored semantically.
The full technical breakdown is available on the related research page.
What changed from the first paper
Our first paper showed that explicit episodic and semantic memory could outperform simpler alternatives, but its policies were still handcrafted. That made the result promising, but incomplete. We still needed to answer a harder question: can an agent learn how to manage explicit memory rather than rely on rules we wrote ourselves?
Explicit Memory answers that question with reinforcement learning in RoomEnv-v1, a more demanding version of our environment. The agent now observes events, stores them in short-term memory, and learns which items should be routed into episodic or semantic long-term memory to maximize future reward.
Why we kept the memory explicit
The memory stays explicit and inspectable and is still trainable. We do not have to choose between symbolic structure and learning: the agent learns the memory-management policy while the memory contents stay separate enough to analyze.
The experiments show that the learned agent can outperform fixed storage baselines, and that pretrained semantic knowledge can improve learning further. So explicit memory does not have to stay handcrafted.
Resources