Agentic AI Interview Questions
Dashboard
0%
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →

Q29IntermediateConcept

Design long-term memory for an agent: what to store, when to write, how to retrieve, and how to forget.

30-second answerSay your answer out loud first, then reveal.
Long-term memory: an extractor LLM turns the conversation into candidate facts, which are reconciled against existing memories to add, update, delete or ignore; a new query retrieves from the memory store by semantic similarity, recency and importance, and the result is injected into context.

Design decisions

  1. What to store: stable preferences, facts about entities, important decisions, outcomes of past tasks, and procedural lessons ("deploy needs VPN first"). Not small talk or secrets.
  2. Granularity: atomic facts ("prefers Python") are easier to update than long summaries. Profile documents (one per user) work well for stable attributes.
  3. When to write:
    ◦ Hot path: immediate, but adds latency and the agent may over-save.
    ◦ Background: better quality and no latency cost, but delayed.
  4. Retrieval: hybrid scoring (semantic similarity + recency + importance). Always filter by user or tenant namespace.
  5. Conflict resolution: "moved from Delhi to Bangalore" must update, not add a contradicting fact. Tools such as Mem0 and LangMem implement add/update/delete reconciliation.
  6. Forgetting: TTLs for time-bound facts, decay, user-requested deletion ("forget X"), and compliance (GDPR right to erasure).
  7. Privacy: don't store sensitive categories without consent; encrypt; audit.

Evaluation: test whether the agent recalls the right fact, doesn't surface stale facts, and doesn't leak one user's memory to another.

PreviousHow do you reduce an agent's cost?

This is what real progress feels like.