Why Long Context Isn’t Enough for AI Memory

The AI industry is fixated on scaling context windows, but Dan Biderman argues that long context alone cannot solve the memory problem. Context rot sets in as conversations and documents pile up, and techniques like RAG and compaction eventually break down under the weight of trillions of tokens. The real tension is between storing information as text and actually learning from it—models that can’t update their own weights are doomed to repeat the same mistakes, no matter how much context they can hold.

Engram‘s approach is to compress knowledge into knowledge cartridges and directly into model weights through continual learning. Instead of dumping everything into a prompt, they use test-time training to update the model’s parameters on the fly, avoiding the prefill destruction that plagues naive approaches. For enterprise use cases like Harvey, this means holistic queries can be answered without stitching together RAG chunks. The bet is that token efficiency is inseparable from intelligence—a model that has truly internalized knowledge uses fewer tokens to answer harder questions.

The takeaway for builders is that persistent, personal AI memory requires a shift from stateless inference to stateful learning. Biderman compares personal AI models to Tamagotchis: they improve over time through user-specific feedback loops, but the infrastructure to support millions of continuously updated models is still nascent. Serious teams should invest in research around continual learning and weight-based memory, because the ability to compress experience into model intuition is what separates a useful assistant from a glorified search engine.