Storing less data to better match what users actually want

Running personal AI agents locally requires storing user context in tight memory. EPIC solves this by focusing on user preferences rather than raw data, filtering what gets indexed and aligning retrieval to match what users actually care about. Across four benchmarks, it cuts memory needs by 2,404× and boosts preference-following accuracy by 20 percentage points, while keeping latency at 29 ms per query on devices.