v0.5.0 — Offline / local Ollama models

Engram now runs fully offline against a local Ollama daemon — semantic search and the LLM judge / HyDE / consolidation with no API keys.

New features

• Local Ollama embedding provider. Set [embedding] provider = "ollama" (e.g. model = "qwen3-embedding:0.6b" at 1024-dim) to produce real dense vectors against a local Ollama daemon with no API key. Switching an existing database is a one-time engram reembed; the [6020] guard enforces that stored vectors match the active provider.
• Local Ollama text generator. Set [llm] provider = "ollama" to run the LLM judge, HyDE, and consolidation on a local chat model (e.g. qwen3:4b) with no API key. An unreachable daemon falls back to the heuristic judge.
• Ollama endpoint configuration — new [embedding].host / [llm].host config fields and the ENGRAM_OLLAMA_HOST environment override select the Ollama host (default http://localhost:11434).
• Ollama in the engram init TUI wizard — selectable as both an embedding and an LLM provider, with no API-key prompt.

Improvements

• Search degrades to FTS5 when embeddings are unavailable instead of failing the request. memory_search returns a { results, degraded } envelope on both the healthy and degraded paths; degraded: true signals the vector branch was skipped (e.g. the embedding provider is unreachable).
• Localhost-tuned retry backoff for the Ollama HTTP clients — short, few retries — instead of reusing the cloud-latency budget. Provider documentation is pinned against the bundled config template and README by a docs-invariant test.

Breaking changes

• None.

---

Install: cargo install engram-memory · npm i -g @engramm/engram-mcp-server · pip install engram-trainer
Full changelog: v0.4.0...v0.5.0