Chatbot Personalization / Memory System Design

MAI / Copilot onsite system design: design the long-term memory and personalization layer behind a text chatbot — how to store, retrieve, and reconcile per-user history so responses are personalized without re-reading everything each turn. Requirements Design the personalization / memory subsystem for a text-only chatbot. Core functional requirements probed in the round: Persist per-user history and salient facts across sessions. On each turn, retrieve the relevant slice of a user's history to condition the response — a RAG-style retrieve-then-generate flow is the expected baseline. Reconcile conflicting information in a user's history: when stored facts contradict, decide what to surface. Bound the cost of growing history so retrieval stays cheap as a user accumulates months of conversation. Incorporate a feedback loop so the memory improves from signal on whether personalization helped. The interviewer drills each of these as a follow-up rather than expecting a full architecture up front. Notes The follow-ups candidates were pushed on: Conflict resolution — given contradictory facts in a user's history, how do you choose which to retrieve: recency, source confidence, explicit user correction, or a reconciliation pass. Latency — retrieving and re-embedding on every generation is expensive; the interviewer steers toward maintaining a pre-generated per-user summary (a rolling profile) read directly, with full retrieval reserved for the cases the summary does not cover. History overflow — when per-user history grows unbounded, summarize / compress older turns and keep an embedding index over the rest rather than retrieving raw transcripts. Feedback loop — close the loop on whether the personalized response landed and feed that back into what gets stored or up-weighted. No hard scale numbers are given; treat it as an open ML system-design round where stating your own assumptions (users, turns per user, embedding dimension, retrieval budget) is part of the signal. Preparation Sketch a RAG memory architecture end-to-end: write path (extract + embed + store salient facts), read path (retrieve top-k + rolling summary), and the generation-prompt assembly. Prepare a crisp answer for conflict reconciliation and for the summary-vs-full-retrieval latency trade-off — the two follow-ups every candidate hit. Be ready to bound history growth with hierarchical summarization and to describe the feedback signal you would log.