HumemAI iconHumemAI
View on GitHubGet early access

Projects / Machines With Human-Like Memory

Subproject

Explicit Memory

The reinforcement-learning implementation behind A Machine with Short-Term, Episodic, and Semantic Memory Systems.

View GitHubRead paper

RL memory

Learn memory management instead of hand-coding it.

Figure 1

Learned memory management

  • Short-term memory
  • Episodic memory
  • Semantic memory
  • Learned memory-management behavior
The repository trains policies over explicit memory systems rather than hiding memory inside a single recurrent state.

Explicit Memory extends the earlier heuristic work by training agents to decide what to retain, move, and forget across short-term, episodic, and semantic memory systems. The goal is not just better performance, but a learnable memory architecture that remains structurally explicit.

That shifts the project from manually specified memory rules to reinforcement learning over memory operations while keeping the memory system itself visible enough to analyze.

RoomEnv-v1

Move the benchmark forward with a richer memory setting.

Figure 2

Training setting

  • Reinforcement learning in partial observability
  • Structured memory actions
  • Knowledge-graph-oriented memory representation
  • Analysis of learned memory behavior
RoomEnv-v1 supports learned explicit-memory agents in a benchmark where memory decisions affect downstream behavior.

The implementation is built around RoomEnv-v1, where reinforcement learning agents operate with explicit memory structures rather than opaque hidden-state memory. This makes it possible to test whether the agent can learn to use structured memory well under partial observability.

The environment and training setup turn memory control into part of the policy-learning problem instead of treating it as a fixed subsystem.

Bridge

Bridge early heuristic memory work to later graph-memory systems.

Figure 3

Research bridge

  • From heuristics to learning
  • Keeps memory explicit
  • Intermediate step in the research line
  • Groundwork for later graph-memory systems
The project connects the first explicit-memory agents to the later temporal knowledge-graph direction.

Explicit Memory sits in the middle of the broader research line. It keeps the commitment to explicit memory systems, but moves from handcrafted agents toward learned behavior.

That makes it the connective step between the first human-like memory systems paper and the later temporal knowledge-graph work.

Resources

Paper and code.

View GitHubOpenRead paperOpenParent projectOpen