Key Concepts
- 1Agent memory splits into short-term (this conversation's working context) and long-term (persistent facts about the user across sessions), and they need different architectures
- 2Short-term memory lives in or near the context window and manages the current session's constraints, decisions, and state — it's about coherence within a conversation
- 3Long-term memory is a persistent store of durable facts (preferences, sizes, past purchases) retrieved into context when relevant — it's about continuity across conversations
- 4Not everything should be remembered long-term: ephemeral or sensitive details, and stale preferences, can harm more than help, so memory needs curation and decay
- 5Long-term memory is personal data subject to consent, retention, and deletion — the architecture must support erasure and respect privacy boundaries
- 6Retrieval into context is the bridge: long-term memory is only useful if the right facts are surfaced at the right moment without bloating the window
- 7Memory quality is a product lever: good long-term memory makes the agent feel like it knows the user, driving retention; bad memory (wrong or creepy) destroys trust
Why It Matters for PMs
Continuity drives retention: Long-term memory makes the agent feel like it knows the user, a key CLV driver. Its architecture directly shapes the relationship.
Curation prevents harm: Remembering the wrong or stale things damages trust. Memory needs decay and curation, not indiscriminate persistence.
Privacy by design: Long-term memory is personal data. Supporting consent, retention limits, and deletion must be architected in, not bolted on.
Test Your Understanding
Question 1 (Multiple Choice)
What is the difference between short-term and long-term agent memory?
Question 2 (Multiple Choice)
Why shouldn't everything be stored in long-term memory?
Question 3 (Multiple Choice)
What makes long-term memory actually useful in a conversation?
Question 4 (Short Answer)
How would you decide what belongs in long-term memory vs short-term working context?
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
What privacy and curation controls would you build into long-term agent memory, and why?
Think about your answer before revealing the explanation.