How it works

From a chat to memory that gets better.

Four steps, all on your machine. Plain files you can read, edit, diff and delete.

capture

Your agent works.

Hooks quietly capture each session and the facts your agent learns, saved as short Markdown notes in a folder you own.

Recall finds it.

Keyword and semantic search together. Ask for “login broken” and get the note that says “auth failed”, inside your token budget.

A dream refines it.

Overnight, a local model merges duplicates, flags stale facts and spots conflicts. It only proposes. You run vexon apply.

Any agent picks it up.

The next session starts informed, in Claude Code, Codex, Cursor or any MCP client. Start in one, finish in another.