How AI assistants manage context
What “memory” means in product UX — and where consumer chat patterns break in regulated work.
Product
Consumer and workplace AI assistants (ChatGPT-class, copilots) — conversation as the interface to retrieval and generation.
Problem solved
Users need continuity across turns without re-pasting the entire world every time — and help that feels personal without becoming a liability.
UX decisions
- Thread as context containerSimple mental model; weak for multi-matter professional work.
- Memory as optional preferenceUseful for tone; dangerous when treated as ground truth.
- Citations when availableBest products show provenance; many still feel authoritative without sources.
Technical observations
Context windows, RAG, and summarization are engineering tactics. The product decision is what is allowed to persist and what requires a human gate before it touches a system of record.
Lessons for builders
For LegalTech: retrieve → draft → cite → human verify. Never silent writes to matter files. Pattern: Human + AI Review. Prototype: playground. Research: document intelligence.