How AI Learns to Forget
Cognitive memory architecture - Ebbinghaus decay, flashbulb memories, and LLM evaluation.
The Problem with Perfect Memory
AI systems have perfect memory. They never forget. They recall every detail with complete accuracy.
This is a disaster for believable social simulation.
Real players do not have perfect recall. They vaguely remember someone was acting suspiciously but cannot recall exactly when. They conflate conversations from different days. They feel certain about events that occurred differently than they remember.
This imperfect memory isn't a flaw - it's essential to authentic social interaction. When an AI cites facts perfectly across hundreds of exchanges, it feels artificial. When it expresses genuine uncertainty because memories have degraded, it feels human. This is crucial for simulating The Traitors convincingly.
What Memory Research Tells Us
Memory is Reconstructive
When you remember something, you're not playing back a recording. You're rebuilding the memory from fragments, filling gaps with plausible details. This introduces errors - and affects voting pattern accuracy.
Emotion Shapes Encoding
Events triggering strong emotions are remembered more vividly and persist longer. The emotion engine models this encoding process.
Forgetting Follows a Curve
Hermann Ebbinghaus discovered memory decays predictably: rapid initial forgetting, then gradual stabilisation. This "forgetting curve" has been replicated countless times.
The Forgetting Curve in Practice
Memory Retention Over Time
Clarity Markers: Communicating Uncertainty
Based on memory strength, each memory is tagged with a clarity marker that guides how the AI expresses certainty:
Flashbulb Memories
Certain moments are encoded with such intensity they persist almost indefinitely. In The Traitors, these include:
- Being directly accused at the Round Table
- Learning a trusted ally was murdered
- Witnessing a dramatic banishment
- Receiving a recruitment offer from the Traitors - especially significant for Green Cloaks
These memories start at maximum strength, decay slowly, and preserve emotional context. The AI can describe these with vivid, persistent detail even many game phases later.
Testing Language Models
I tested three language models to see which best expressed appropriate uncertainty from memory markers:
GPT-4o-mini
3.1s response
Excellent emotional integration
Qwen 7B
9.3s response
Good but emotionally flat
Qwen 14B
53.6s response
Too slow for practical use
The Emotional Gap
The most significant difference emerged in emotional integration. For high anxiety (0.75):
GPT-4o-mini:
"I... I don't know, honestly. Marcus was the last person seen leaving before the poisoning, and I noticed he was whispering with Sarah. It's hard to shake off that feeling, but I'm not sure if we should jump to conclusions."
Qwen 7B:
"I'm not sure what to make of Marcus. I saw him whispering with Sarah before the meeting. I guess we'll have to wait and see."
For anxiety of 0.75, "wait and see" is far too passive. The hesitation, uncertainty markers, and emotional acknowledgement in the first response create authenticity the second lacks.
Key Takeaways
- Perfect memory is a liability for believable social simulation - designed forgetting creates authenticity. This applies to all player archetypes
- Memory quality markers (clear, fuzzy, vague) enable appropriate uncertainty expression
- Emotional encoding prioritises significant events, mirroring human memory enhancement
- Source monitoring enables appropriate attribution and hedging in dialogue
- Language model selection dramatically affects emotional expression quality - key for the overall AI architecture
- Local models suffice for development; production benefits from more capable cloud models. Future research will continue to explore model selection