Support local Ollama embeddings + LLM provider (cost-free, private, no external API dependency)

[imonroe]

Summary

Add first-class config support for running mem0's embedder and/or fact-extraction LLM against a local Ollama instance, so memserv can run with zero per-call cost and no data leaving the host.

Idea from OB1

OB1's recipes/local-ollama-embeddings provides "on-device embedding generation" as an alternative to cloud embedding APIs — cutting cost and keeping content private.

Why it fits memserv

• app/memory.py::_build_config already abstracts embedder.provider and llm.provider from settings — mem0 supports Ollama for both. This is largely a config + docs feature, not a re-architecture.
• Directly addresses two PRD §19 operational footguns: embedding cost and the Anthropic fact-extraction bill for high-volume automation.
• Privacy win for a self-hosted single-user system: no content sent to OpenAI/Anthropic.

Proposed approach

• Extend Settings (app/config.py) to accept mem0_embed_provider=ollama / mem0_llm_provider=ollama plus OLLAMA_BASE_URL and model names.
• Wire the Ollama-specific config keys through _build_config.
• Critical reminder: MEM0_EMBED_DIMS must match the Ollama embed model's real output dimension, and changing models requires dropping/recreating the Qdrant collection (per the existing architecture invariant). Document this prominently.
• Add docker-compose.yml notes / an optional Ollama service for the bundled-stack deploy path.
• Document in docs/USER_GUIDE.md.

Notes / scope

Keep the default as anthropic+openai; Ollama is opt-in via env. No changes to the single-Memory-per-process invariant.

Source: https://github.com/NateBJones-Projects/OB1/tree/main/recipes