Meta releases 'Lighthouse-2' API with context windows tailored for UX sessions

AI · 4 min read

Lighthouse-2 increases context retention by blending local session memory with aggregated summaries, enabling the model to recall user interviews, testing scripts, and iteration notes across weeks. The model supports multi-user role tokens, so designers, PMs, and researchers can query the same assistant with preserved role contexts.

The API includes built-in transcription alignment, timestamped references, and a 'session playback' feature that replays conversation highlights and model reasoning. Meta emphasized low-latency performance for live workshops and integrations with popular remote research platforms.

Privacy controls are central: teams can set retention policies per session and toggle anonymization that strips PII before it enters aggregated memory. Meta plans to offer on-prem options for enterprise clients with stricter compliance needs.

Some researchers raised concerns about storing sensitive interview data in an external API; Meta responded by publishing its evaluation of data minimization practices and third-party audits planned for Q3.