Skip to content

v0.6.8

Choose a tag to compare

View all tags
@github-actions github-actions released this 13 Oct 20:16
· 37 commits to main since this release
v0.6.8
4e1aa4c

The desktop app is updated too: Desktop v0.0.2.

Atomic Chat Core is the local engine

  • Local models now run through Atomic Chat Core, the engine the Atomic Chat app uses. It downloads and manages llama.cpp for you. Existing setups move over on first start (config v77). Your model files are kept and nothing is downloaded again.
  • The engine choice is no longer shown in the TUI. atomic-agent models engine llama-server switches back if you need to.
  • Context size and GPU placement are fitted to your machine by Core.

Memory search by meaning

  • New embedding model: EmbeddingGemma 2 from Google (310 MB, 100+ languages). Turn it on with atomic-agent models pull-embedding embeddinggemma-2, models use-embedding embeddinggemma-2 and models start-embedding. It runs through Atomic Chat Core.
  • Memory now finds notes by meaning, not only by shared words. A note in one language can be found by a question in another.
  • Notes written before you turned embeddings on, or before you switched models, are re-embedded in the background after start. atomic-agent memory reindex does it by hand.
  • Fixed: semantic search turned itself off after 200 notes, and keyword ranking in hybrid search was inverted.
  • memory.embeddings.fts5Weight, vectorWeight and bruteForceCeiling in config.json now take effect.
  • Semantic search works while a local chat model is active. With a cloud chat model, memory uses keyword search for now.

Projects and cloud models

  • Cloud workspaces load the project's own instructions (AGENTS.md) and skills. Outside a git project only the folder root is read, with a size limit, so starting in your home folder no longer overflows the model.
  • Cloud models keep the full conversation context, and provider policies are migrated.

Local models

  • Faster model switching.
  • Add a GGUF you already have by its path: in the Local models panel press a and paste the file path. Nothing is copied or downloaded.
  • An unusable llama-server.key is regenerated instead of running without a key.

Fixes

  • Piped command output is no longer cut at 64 KB. Lists such as models search --json arrive whole.
  • A rate limit is no longer read as a billing problem, and a moderation refusal no longer as a bad API key.
  • Fusion: a role's model is reset when its provider is removed or changed, so the old model no longer sticks.
  • Finished turns show a one-line summary under their tool cards in the TUI.
Assets 10
Loading