Prediction: we’re on the cusp of dramatic improvements in AI utility through episodic memory systems. While we’ve made remarkable progress in model capabilities, the ability to maintain coherent, long-term memory across conversations remains a significant bottleneck that will define the next phase of AI development.
OpenAI’s recent “improved” memory feature for ChatGPT represents an early step in this direction, though user feedback has been mixed. This highlights the complexity of the challenge—memory isn’t just about storing information, but about intelligently retrieving and applying it across extended interactions.
Anthropic’s multi-agent research system[1] offers a glimpse into more sophisticated approaches. Their system uses memory to persist research plans when context windows approach the 200k token limit, preventing critical information loss during complex, multi-step investigations. This isn’t just about storage—it’s about maintaining continuity and coherence across extended reasoning processes.
What’s striking is how few frontier labs seem to be aggressively pursuing memory solutions, despite having access to vast external data stores. Google and Meta, for instance, sit on troves of user data that could enable rich, personalized memory systems. The opportunity cost of not developing these capabilities is becoming increasingly apparent. Memory at this point seems to be a footnote more than a focus.
Projects like Graphiti[2] are already exploring this space, demonstrating that the open-source community recognizes the importance of this challenge.
The implications extend far beyond chat interfaces. If this AI summer is going to yield AGI/ASI or something like it, it will likely be by combining LLMs with multiple subsystems—like memory and tool use—much like how human cognition is only one part of what makes us us. Memory systems that can maintain context across months of interactions, build genuine relationships with users, and provide truly personalized experiences represent a fundamental shift in how we interact with AI systems.
The companies that solve memory effectively will have a significant competitive advantage. The question isn’t whether this will happen, but which organizations will lead the charge.
[1]: Anthropic — How we built a multi-agent research system [2]: Graphiti on GitHub