Leon AI 2.0 overview: memory and llama.cpp support

Uploaded on March 18, 2026, this development update covers work toward Leon AI 2.0. The creator describes a future developer preview, desktop installer and smoother onboarding; the video does not announce a completed 2.0 release.

Memory is a central topic. The creator says Leon stores conversation fragments in SQLite and Markdown, then retrieves them through full text search and vector similarity. A demonstration recalls a conversation about baby care from roughly three weeks earlier and identifies a related file. Recall after a full year remains untested, according to the speaker.

For local LLM use, the creator discusses a recent llama.cpp implementation, chosen after initially trying Ollama. Running the main model locally depends on sufficient hardware. Leon also supports hosted model providers, with differences in their APIs and reasoning controls requiring adjustments to the agent loop.

The update describes context that loads when relevant tools need it, a reflection diary and a periodic pulse for pending tasks. In one example, Leon initiates a runtime check that the creator interprets as help with earlier provider-switch problems. The report recalls crashes and context-limit trouble, so this remains a development demonstration rather than proof of reliability. Access to computer habits and browser history requires user permission.