Changelog

Every release, most recent first. Dates are when the change shipped, not when it was written.

2026-07-28 · Cloud replica, first working build

Unreleased

• Read-only cloud replica: Markdown to object storage, metadata and full-text search to a per-tenant database, embeddings to a vector index.
• Remote MCP server over Streamable HTTP with OAuth 2.1, PKCE, and dynamic client registration — connects to Claude, ChatGPT, Gemini, and Cursor from one endpoint.
• Four read tools: memory_search, memory_get, memory_briefing, memory_timeline.
• Desktop sync agent with incremental push, deletion detection, and a --inspect mode that reports vault problems before sending anything.
• Nightly sweep of expired authorisation codes and tokens.

2026-07-27 · Decisions that shaped the build

Unreleased

• The cloud re-embeds rather than accepting vectors from the desktop. The two run different embedding models, and vectors from different models are not comparable even at identical dimensions — a mismatch would load without error and quietly return nonsense.
• Version one is read-only. With a single writer there is no merge to perform, which removes conflict resolution — the hardest part of the architecture — from the first release entirely.
• The replica is readable by our servers. Searchable-while-offline and zero-knowledge cannot both be true today. Documented on the security page rather than glossed over.

Earlier

Varven grew out of a personal memory system: an Obsidian vault of Markdown notes, a local model doing embeddings and summarising, and a protocol server exposing it to a desktop AI client. It worked for one person on one machine for months before any of this existed.