Are you a founder actively fundraising? Apply to be featured in the Lynx List
Join us on Aug 6 for Building & Investing in AI: Insights from OpenAI + AI Investors
Daniel Porras Reyes is an investor at Flybridge Capital, a seed-stage venture firm focused on AI. In a recent piece titled Memory the Missing Piece for Real Intelligence (also published on his Substack), he explains why memory is a critical component for building more intelligent, adaptive, and valuable AI applications.
Memory is one of the most critical and underdeveloped components of the modern AI stack. While large models have made major leaps in reasoning and generation, they still suffer from what he calls “functional amnesia.” Below are the key takeaways from his piece:
1. Larger context windows are not enough
Even as models are trained with billions of parameters and context windows now span millions of tokens, most AI agents still forget what happened minutes earlier. This limits their ability to learn from past interactions or develop persistent behavior over time.
2. True intelligence requires memory
For AI agents to act more like humans, they must be able to remember relevant facts, experiences, and user preferences. Without memory, agents are static tools that must restart every session without continuity.
3. Memory is a new layer in the AI stack
Just like we’ve seen layers emerge for model inference, deployment, and fine-tuning, memory infrastructure is forming its own category. This includes systems to select important information, store it efficiently, and recall it when relevant.
4. The ecosystem is starting to take shape
Startups are building tools across this layer, including vector databases, memory APIs, retrieval-augmented generation pipelines, and agent frameworks with built-in memory. These tools are aimed at helping agents better retain and reuse useful information.
5. Strategic questions remain open
There are still important question to be answered: where will value accrue in the memory layer? Will it belong to foundation model providers, to developers of memory frameworks, or to application-level products that best leverage memory?
6. The opportunity is wide open
Memory remains one of the least mature yet most essential areas in AI infrastructure. For founders and investors, this represents a wide open space for innovation—with implications for personalization, long-term task execution, and the next generation of intelligent agents.
Share this post