Memory Systems

An AGI system cannot rely on a static weight matrix alone. It needs a dynamic, multi-layered memory system to store experiences, facts, and skills.

The Multi-Layer Memory Model

AGI memory is often modeled after human cognitive psychology, distinguishing between different durations and types of information.

1. Short-Term / Working Memory (STM)

The "scratchpad" of the mind. It stores information currently being processed.

• In AGI: Attention windows, current state buffers, or "Active Atoms" in the AtomSpace.
• Capacity: Limited but highly accessible.

2. Long-Term Memory (LTM)

Persistent storage of knowledge and experience.

Visualizing Memory Flow

graph TD
    SensoryInput[Sensory Input] -->|Attention| WM[Working Memory]
    WM -->|Encoding| LTM[Long-Term Memory]
    LTM -->|Retrieval| WM
    
    subgraph LTM Structure
    Semantic[Semantic Memory: Facts]
    Episodic[Episodic Memory: Events]
    Procedural[Procedural Memory: Skills]
    end
    
    Encoding --> Semantic
    Encoding --> Episodic
    Encoding --> Procedural

Forgetting and Consolidation

A crucial part of AGI memory is Active Selection. If a system remembers everything perfectly, it eventually becomes overwhelmed by noise.

• Consolidation: Moving important working memory items into LTM during "down-time" (similar to sleep).
• Graceful Forgetting: Lowering the Importance (STI/LTI) of rarely used Atoms until they are purged or archived.

Implementation in Hyperon

In the Hyperon framework, memory is unified in the AtomSpace.

• Every Atom has Attention Values (STI/LTI).
• This allows the system to treat memory as a dynamic, self-organizing graph rather than a static database.

Next: Reasoning Engines