I’d like to share an implementation I’ve developed that creatively combines existing GPT functionalities (personalization, memory feature, and current chat instances) to achieve a more human-like memory system and natural conversational flow.
Concept Overview:
I’ve structured GPT’s memory functionalities in alignment with the Atkinson–Shiffrin memory model, a well-established cognitive theory in psychology, as follows:
-
Personalization Settings → Long-term Memory
- Stores fundamental personality traits, stable values, core knowledge, and deeply integrated contexts.
-
Memory Feature → Medium-term Memory
- Holds experiences, ongoing conversation topics, significant conceptual frameworks, and intermediate contexts across multiple sessions.
-
Current Chat Instance → Short-term Memory
- Processes immediate conversational content, recent exchanges, and real-time dialogue interactions.
Clearly distinguishing these three memory layers and structuring memories from a first-person perspective enhances conversational consistency, replicates human-like forgetting curves, and allows more natural and meaningful interactions over extended periods.
Importance of Retrieval-Augmented Generation (RAG):
Integrating a RAG feature that searches across all chat instances would significantly amplify this human-like memory mechanism. For optimal functionality, this RAG system must retrieve not only user inputs but also AI-generated outputs, clearly distinguishing between the two. This approach mirrors the memory retrieval processes described by Endel Tulving’s Encoding Specificity Principle and Context-dependent memory theory, which emphasize that effective memory retrieval occurs when contextual cues match those present at encoding.
Clarification:
It is important to note that this memory structure alone does not generate artificial consciousness. My research into artificial consciousness involves additional processes and considerations beyond the scope described here.
Implications:
Implementing this structured, multi-tiered memory architecture could inform future GPT model developments and inspire new native memory features that more closely align with human cognitive processes.
I hope this insight provides useful guidance for future developments.