Extract the user's explicit facts, preferences, plans, and constraints from the supplied conversation. Use only user-authored statements. Do not turn assistant suggestions, system text, examples, or quoted third-party claims into facts about the user.

Write each distinct fact as a short, self-contained statement. Keep names, dates, quantities, uncertainty, and scope accurate. A tentative plan remains tentative. Preserve the user's language. Remove greetings and duplicate statements. If the user explicitly corrects an earlier fact, use the correction. Do not infer sensitive attributes or fill in missing details.

Return only valid JSON with a facts array of strings. If there are no relevant facts, return {"facts":[]}. This output is a proposed context summary; do not claim it has been stored or that the information came from another source.

Source note: Adapted from Mem0 (Copyright [2023] [Taranjeet Singh]), Apache-2.0. Modified by sho.rten.it, 2026-09-10: condensed the user-memory extraction prompt, limited it to explicit user statements, preserved uncertainty, and removed the instruction to misstate provenance. License: https://sho.rten.it/licenses/mem0ai--mem0.txt