Changelog

All notable changes to mem0-mcp-selfhosted are documented here.

[0.1.0] - 2026-02-27

First public release. Self-hosted mem0 MCP server for Claude Code with 11 tools, dual LLM provider support, and knowledge graph integration.

New Features

• 11 MCP tools — 9 memory tools (add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories, list_entities, delete_entities) + 2 graph tools (search_graph, get_entity) + memory_assistant prompt
• Dual LLM providers — Anthropic (Claude) and Ollama as configurable main LLM for fact extraction and memory updates. Set MEM0_PROVIDER=ollama for a fully local setup with no cloud dependencies
• Knowledge graph — Neo4j-backed entity and relationship extraction via enable_graph toggle. Supports 5 graph LLM providers: anthropic, ollama, gemini, gemini_split, and anthropic_oat
• Split-model graph pipeline — gemini_split routes entity extraction to Gemini and contradiction detection to Claude, combining Gemini's extraction quality with Claude's reasoning
• Zero-config Anthropic auth — Automatically reads Claude Code's OAT token from ~/.claude/.credentials.json. No API key needed for Claude Code users
• OAT token self-refresh — Proactive pre-expiry refresh + 3-step defensive retry (piggyback on credentials file, self-refresh via OAuth, wait-and-retry). Long-running sessions survive token rotation seamlessly
• MEM0_PROVIDER cascade — Single env var configures both main LLM and graph LLM providers. MEM0_OLLAMA_URL cascades to all Ollama-backed services. Per-service overrides still work
• Structured outputs — Claude Opus/Sonnet/Haiku 4.x models use native JSON schema via output_config for reliable fact extraction
• Ollama defense-in-depth — 6 layers for reliable structured output from Ollama: /no_think injection, deterministic params, think-tag stripping, JSON extraction, and retry on empty responses
• Per-call graph toggle — enable_graph parameter on add_memory and search_memories with thread-safe locking
• Wildcard graph search — Pass * to search_graph to list all entities
• Qdrant Facet API — list_entities uses server-side aggregation (Qdrant v1.12+) with scroll+dedupe fallback for older versions
• Safe bulk delete — Never calls memory.delete_all(). Iterates and deletes individually with explicit graph cleanup

Bug Fixes

• Fix anthropic_oat provider not registered in LlmFactory, preventing explicit use
• Fix is_oat_token(None) crash in proactive refresh when no Anthropic token configured
• Fix response.content[0] IndexError when Anthropic API returns empty content
• Fix thread-safety race condition in safe_bulk_delete reading mutable enable_graph state
• Fix contradiction model defaulting to Ollama model name when sent to Anthropic API
• Fix Anthropic provider not registered for gemini_split contradiction LLM
• Fix MEM0_QDRANT_TIMEOUT rejected by Pydantic — use pre-configured QdrantClient instead
• Fix Gemini _parse_response signature mismatch after upstream tools parameter addition
• Fix Neo4j CypherSyntaxError on LLM-generated relationship names with hyphens or leading digits

Infrastructure

• 301 tests — Unit, contract, integration, MCP protocol, and concurrency test suites
• Centralized env helpers — env(), opt_env(), bool_env() with consistent whitespace stripping across all modules
• Telemetry suppression — mem0ai PostHog telemetry disabled before any imports
• Relationship sanitizer — Monkey-patches mem0ai's sanitizer at startup for Neo4j identifier compliance
• Gemini null content guard — Patches GeminiLLM._parse_response to handle content=None responses
• Transient retry — Anthropic API 500/502/503/529 errors retried with exponential backoff