Myths, facts, and what to keep
Lesson 5 of 5 in Agent Memory: Remembering and Forgetting on Purpose.
Memory attracts more folklore than any other part of agent design, because the product experience — "it remembered me!" — feels so much like cognition. Each myth below sounds reasonable, ships in real products, and fails in a specific, predictable way.
Myth: “The model remembers our conversation”
The model is stateless — a pure function from context to output, retaining nothing between requests. Within a session, the application re-sends the history every turn. Across sessions, any "remembering" is an engineered pipeline that extracted facts into a store someone built, someone maintains, and someone should be able to show you. When a product remembers you, the correct question is never "how smart!" — it is "where is that stored, and can I read it?"
Myth: “Bigger context windows make memory obsolete”
A context window, however large, is per-request capacity, not persistence — it still evaporates when the session ends. And filling it is not free: cost and latency scale with every token you stuff in, and models attend unevenly across very long contexts, so the fact you buried at token 400,000 may as well be forgotten. Long-term memory solves a different problem than capacity: surviving the end of the conversation.
Myth: “More memory always makes the agent better”
Memory is not a score to maximize. Stale entries cause confident wrong actions; irrelevant entries dilute the model’s attention on the task; and every stored fact is token cost, attack surface, and privacy liability. The best memory systems are curated, not hoarded — they forget aggressively and remember deliberately. Lesson 4's write gate exists precisely because the default answer to "remember this?" should be no.
Myth: “The vector database is the memory system”
Storage is the easy half. A memory system is the set of policies: what qualifies for writing, whose words get trusted, how facts expire, what the user can see, and how a poisoned entry gets found and purged. A vector store without write gates and provenance is not a memory system — it is a poisoning target with cosine similarity.
Myth: “Deleting the chat deletes what the agent knows about me”
Only if nothing was derived from it. If a memory pipeline extracted facts from the conversation into a long-term store — and possibly into embeddings, caches, and backups of that store — the transcript’s deletion orphans those copies rather than removing them. Real deletion is an architecture property: every derived artifact must be traceable back to its source so the delete can chase it. That trace is provenance again, doing double duty.
Interactive flashcard deck.
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.