Tempera

Gives Claude Desktop persistent memory across coding sessions through semantic episode capture and retrieval. Instead of forgetting everything between chats, Claude can store what worked, search past solutions, and learn from feedback through a reinforcement learning loop that boosts helpful memories and fades stale ones. Exposes tempera_capture, tempera_retrieve, and tempera_brief tools. The brief surface loads pending corrections, calibration warnings, and relevant patterns before you start a task. Dream cycle runs nightly to detect contradictions and extract reusable templates. Supports cross-project learning with scoped decay rates so universal truths persist while project workarounds expire naturally. Built in Rust with local BGE embeddings, no cloud dependencies.