Meta AI releases Llama Dialogue 2 with improved persona memory and safety filters

AI · 5 min read

Llama Dialogue 2 focuses on persistent persona memory and multi-turn coherence to support assistant-style integrations in creative tools. Meta improved memory primitives so the model can store and recall style preferences, project constraints, and recurring user corrections across sessions.

Safety enhancements include rate-limited memory writes and reviewable summaries of remembered facts so users retain control over what the model stores. The model's system also employs layered moderation and content-scoring to reduce harmful outputs in creative content generation.

Meta provided developer tooling for memory gating, allowing apps to request or deny memory access based on workspace context. Several creative app developers have begun testing the model in closed beta to power long-running assistant features.