mem0-api-server

Unified AI memory + document RAG API for the aienv stack. Wraps:

• Qdrant (vector embeddings, semantic search)
• OpenSearch (BM25 full-text)
• Neo4j (entity relationship graph)
• PostgreSQL (relational metadata)

Hybrid search (semantic + BM25 with RRF merge), persistent memory via the mem0ai library, and an OpenAI-compatible /v1/chat/completions endpoint that injects RAG context before forwarding to Ollama.

Endpoints

Path	Purpose
GET /health	Service + dependency health
POST /v1/memories/	Add a memory
POST /v1/memories/search/	Semantic memory search
GET /v1/memories/?user_id=...	List memories
POST /v1/conversations/	Ingest a conversation transcript
POST /v1/conversations/search/	Search conversation history
POST /v1/documents/index	Index a document (chunks + embeds + indexes in OS)
POST /v1/documents/query	Hybrid search (Qdrant + BM25 → RRF)
POST /v1/documents/query-llm	RAG query → LLM answer with citations
GET /v1/documents/collections	List document collections
POST /v1/chat/completions	OpenAI-compatible chat with implicit RAG injection
POST /mcp/	Streamable HTTP MCP server (rag_query, mem0_search, ...)

Deploy

Container image is published to git.loop-coop.net/loco/workloads-mem0-api-server:vX.Y.Z AND :latest by Woodpecker on every tag push. Consumers (the aienv-mem0 qubes module) pull this image with Pull=newer — no local builds, no version pinning. A new release lands by tagging this repo and restarting the consumer:

edit + commit + push to main
  → Woodpecker bump.yml auto-tags vX.Y.Z (Conventional Commits)
  → release.yml builds + pushes :vX.Y.Z + :latest
mem0 qube:
  → systemctl restart mem0
  → Pull=newer fetches the new digest, container starts on it

No aienv-salt --sync, no apply mem0 for routine releases.

Runtime config is a YAML mounted at /app/config.yaml (Qdrant/OpenSearch/Neo4j hosts + ports, embedder/LLM model). See config.yaml.example.

The service is agent-neutral — MCP tools default user_id to "default". Each MCP client (Claude Code, opencode, custom agents) passes its own user_id so memories segment per agent.

Tools (MCP)

The /mcp/ endpoint exposes Streamable HTTP MCP tools so any MCP-aware client (Claude Code, opencode, custom agents) can scope RAG queries and read/write mem0 directly. The server is agent-neutral — no client is the default; identity is whatever the caller passes as user_id. Six tools: rag_query, rag_query_llm, rag_list_collections, mem0_search, mem0_add, mem0_list.

Client config:

"rag": {"type": "http", "url": "http://localhost:8090/mcp"}

See docs/MCP.md for the full tool reference, smoke-test commands, implementation notes, and the comparison between explicit /mcp/ calls and the implicit RAG injection in /v1/chat/completions.