HumemAI iconHumemAI
View on GitHubGet early access

Projects / Machines With Human-Like Memory

Subproject

Human-Like Memory Systems

Heuristic agents for A Machine With Human-Like Memory Systems, focused on explicit episodic and semantic memory in RoomEnv-v0.

View GitHubRead paper

Heuristic agents

Start with explicit memory systems before learned policies.

Figure 1

Early explicit memory agents

  • Handcrafted agent behavior
  • Explicit episodic memory
  • Explicit semantic memory
  • Inspectable decision logic
The repository isolates memory-system design with heuristic policies rather than end-to-end training.

This project studies whether human-inspired memory structure helps in a simple partially observable setting before introducing reinforcement learning. The agents are hand-built so episodic and semantic memory behavior stays visible and debuggable.

That makes the repository a clean first step in the wider research line: memory is explicit, decisions can be inspected, and the architecture can be compared directly.

RoomEnv-v0

Test memory retrieval in a controlled partially observable environment.

Figure 2

Evaluation setting

  • Partially observable rooms
  • Delayed questions
  • Stored observations as memory
  • Direct comparison of memory variants
RoomEnv-v0 provides a simple benchmark where memory quality affects downstream question answering.

The implementation is built around RoomEnv-v0, where the agent has to navigate, retain observations, and answer questions later without privileged access to the full world state.

That setup lets the paper ask a narrower question: what do explicit episodic and semantic memory systems contribute when the task genuinely requires remembering.

Starting point

Use this repository as the opening move of the whole project line.

Figure 3

Research sequence

  • Initial architectural framing
  • Baseline for later learned systems
  • Directly tied to the first paper
  • Clear comparison point for later work
The project establishes the first explicit-memory implementation in the broader research sequence.

Human-Like Memory Systems is the earliest implementation step in Machines With Human-Like Memory. It establishes the architectural framing before the work moves on to reinforcement learning and temporal knowledge-graph memory.

That is why the repository still matters on its own: it makes the initial claims concrete and gives the later systems a clear point of comparison.

Resources

Paper and code.

View GitHubOpenRead paperOpenParent projectOpen