Memory Experiments

How does an AGI manage its knowledge? These experiments test the efficiency of the AtomSpace memory model.

Experiment 1: The "Needle in the Haystack"

We populate the AtomSpace with 1,000,000 random facts and 1 specific target fact.

• Goal: Measure the time it takes the Pattern Matcher to find the specific fact.
• Result: Using indexing, retrieval is near-instant, showing the scalability of the hypergraph.

Experiment 2: Concept Generalization

We feed the agent 100 sentences about different "Fruits."

• Goal: Can the agent automatically create a ConceptNode "Fruit" and link the specific items to it?
• Result: Using a simple clustering algorithm on MeTTa atoms, the system successfully categorizes 90% of the items.

Experiment 3: Consolidation (Sleep)

We let an agent act in a simulation, generating 50,000 "Event Atoms."

• Goal: During "Down-time," the agent summarizes these events into 10 key "Lessons Learned."
• Visual Process:

graph TD
    Events[50k Raw Events] --> Filter[Importance Filter]
    Filter --> Summary[Episodic Summary]
    Summary --> LTM[Write to semantic LTM]
    Events -.-> Delete[Purge Noise]

Why these matter

An AGI must be able to Filter Noise from Signal. These experiments show that the Hyperon Attention Value (STI/LTI) system is an effective way to mimic biological memory consolidation.

Next: Notes section