4 lessons
Conversation memory patterns
How chatbots appear to "remember" a conversation when the model itself has no memory at all — the actual patterns systems use to fake continuity.
The math of summarization-based memory
How much do you actually save by summarizing old context, when does it break down, and how to think about the compression ratio precisely.
How context windows get stretched beyond their training length
RoPE scaling, ALiBi, and sliding/sparse attention — the actual architectural tricks that let models handle sequences far longer than anything they saw during training.
Vector memory and long-term recall across sessions
Turning "remembering the user" into a retrieval problem — storing past interactions as searchable vectors and pulling back only what's relevant right now.