Memory - short-term vs. long-term
"Memory" in AI products isn't the model learning - it's engineering. Short-term memory is just re-sending the recent conversation each time (limited by the context window, A4). Long-term memory is a separate system that saves facts about you somewhere (a database) and selectively re-inserts relevant bits into future prompts - which is exactly how the memory you're experiencing in this very conversation works.
Building real long-term memory means deciding: what gets saved, how it gets retrieved (often via embeddings/RAG, C1/C5), and how much gets re-injected without blowing the context window. Badly designed memory either forgets things that matter or stuffs in irrelevant details that confuse the model - a real design problem, not a solved default. ChatGPT's "Memory" and Claude's "memory" features are consumer-facing versions of exactly this system, just built and tuned by the provider instead of you.
Notice, right now, that this conversation "knows" things about your work at Kasper from before - that's long-term memory in action: facts stored elsewhere, silently re-inserted into context, not the model recalling on its own.
How does "long-term memory" in AI products actually work?