AI memory is how sessions remember visitor interests, conversation history, and durable facts without stuffing the context window.
In my daily workflow
- Short-term: sliding window of recent messages in the thread.
- Session: visitor interest profile updated from questions and clicks.
- Long-term: only persisted facts that passed validation and consent.
- I prune aggressively — memory should be signal, not a transcript archive.
How it makes me work smarter
Memory makes follow-up questions feel coherent ('you asked about RAG earlier'). Without layering, every message is day one. With bad memory, you leak stale or wrong facts. I optimize for recency-weighted summaries and explicit user corrections.
My setup
- Client-side visitor interest storage with caps
- Server-side thread history for assistant sessions
- Supabase for durable preferences when authenticated
- Summarization job when threads exceed token budget
On this portfolio
The personalization layer tracks assistant questions and interests to tune homepage emphasis. The assistant thread keeps recent turns for context without sending the entire site corpus each request — that's RAG's job.
