How it works
- Opt-in — requires the Specialist tuning pack plus Settings toggle. Nothing persists without your approval in the modal.
- Scopes —
agent(one expert),global(all your experts), orcollaboration(one collab session). - Retrieval — Ollama embeddings (default
nomic-embed-text) with keyword fallback when offline; top-k injection per turn, not dump-all. - Lifecycle — edit, export/import JSON bundles, optional agent-suggested proposals (still need confirmation).
Where you manage it
- Approval modal with scope toggle when an agent proposes a learning.
- Settings grouped lists and per-agent info with scope badges.
- LoRA train preview can include up to 50 confirmed learnings (
include_learnings=1).
Privacy posture
Learnings are stored under ~/.neural-junkie/learnings.json with per-user isolation in v2. This is complementary to rolling session summaries (context model) — learnings are durable facts you explicitly save; summaries compress recent conversation.