Memory Decay: An Innovative Approach to AI Memory

Memory Decay explores the concept of human-like memory in artificial intelligence, suggesting that perfect recall may not be ideal for AI interactions. This project investigates how incorporating natural forgetting patterns can enhance AI communication, making it more relatable and less unsettling.

Why AI Should Forget

The conventional AI approach treats all memories equally, regardless of their age or relevancy. This project proposes several benefits to implementing a forgetting mechanism:

• Honest Uncertainty: AI can express doubt, stating phrases like "I think you mentioned..." when unsure, leading to more genuine interactions.
• Recency Matters: Recent conversations hold more significance than older ones, prioritizing current and reinforced information for relevant exchanges.
• Natural Conversations: Over time, an AI equipped with this memory model behaves like a human friend, who may need reminders instead of having perfect recall.
• User Control: Users can dictate which memories are maintained or allowed to fade based on their relevance.
• Reduced Uncanniness: Forgetting mimics human behavior and fosters a more comfortable interaction atmosphere.

The Method: Ebbinghaus Curve Implementation

Utilizing Hermann Ebbinghaus's forgetting curve, the project applies a mathematical model to simulate memory decay in AI:

1. Natural Memory Decay Rates: Different types of memories decay at varying rates, such as:

   • Facts: 0.01 (very slow decay)
   • Preferences: 0.05
   • Goals: 0.15
   • Events: 0.25
   • Context: 0.60 (fades quickly)

2. Reinforcement of Memories: Each time a memory is reinforced, its decay slows:

   adjusted_decay_rate = base_rate / (1 + 0.3 × reinforcement_count)
   

3. Confidence via Decay: Over time, confidence decreases based on memory decay:

   confidence = initial_confidence × e^(-decay_rate × hours_elapsed)
   

4. Dynamic Verification Needs: Various confidence thresholds dictate AI behavior:

   • ≥0.7: High confidence - use directly.
   • 0.5-0.7: Medium confidence - present with caution.
   • 0.3-0.5: Low confidence - verify before using.
   • <0.3: Archive or delete.

Interactive Demonstration

To see this system in action, access the interactive visualization:
Try the live demo →
Alternatively, download and open index.html locally to experience the memory dynamics firsthand.

Design Philosophy

This project is rooted in the belief that AI interactions should prioritize:

• Honest signals of uncertainty.
• Recent information that is reinforced.
• Natural human-like conversations.
• User agency in managing memory longevity.

Technical Overview

The decay calculation is performed efficiently, allowing real-time adjustment of memory confidence:

private calculateCurrentConfidence(memory: Memory): number {
  const hoursSinceReinforced =
    (Date.now() - new Date(memory.last_reinforced).getTime()) / (1000 * 60 * 60);

  const adjustedDecayRate =
    memory.decay_rate / (1 + 0.3 * memory.reinforcement_count);

  const confidence =
    memory.initial_confidence * Math.exp(-adjustedDecayRate * hoursSinceReinforced);

  return Math.max(0, Math.min(1, confidence));
}

Memories are tracked using parameters such as initial confidence, decay rate, reinforcement counts, and the last reinforced timestamp. This allows for confident and controlled memory recall.

Conclusion

Memory Decay aims to transform AI interactions, fostering a more human-like dialogue rhythm. Instead of presenting facts with unwarranted confidence, the AI poses questions that validate memory, thereby honoring the natural forgetting process.

Explore the project and witness the evolution of AI memory dynamics.