Mem0

Stores conversational history, user preferences, and domain knowledge across user, agent, and session scopes using LLM-powered fact extraction that automatically identifies and deduplicates relevant information from raw conversation text. The system uses configurable LLM providers (18+ supported) to parse unstructured input into structured memory entries, then persists them across vector stores (24+ backends) and optional graph databases for semantic retrieval and relationship tracking.

Retrieves relevant memories from storage using semantic similarity search powered by configurable embedding providers (11+ supported including OpenAI, Cohere, Ollama) and optional reranking to improve relevance. The system converts query text to embeddings, searches across vector stores with configurable similarity thresholds, and optionally applies cross-encoder reranking to re-score results before returning to the application.

The Platform deployment exposes a REST API with built-in multi-tenancy support through organizations and projects, enabling SaaS applications to manage multiple customers' memories in isolation. The API includes authentication via API keys, organization/project scoping, user management, and webhook support for memory events, allowing external systems to react to memory changes.

Provides native integration with popular AI frameworks through adapters and plugins, including Vercel AI SDK provider integration and OpenClaw plugin support. These integrations allow memory operations to be seamlessly embedded into agent workflows without manual orchestration, with automatic context passing and memory updates.

Enables exporting all memories for a user, agent, or session in multiple formats (JSON, CSV, etc.) for data portability, compliance (GDPR data subject access requests), or migration to other systems. The export operation retrieves all memories matching filter criteria and serializes them in the requested format with full metadata and audit trail information.

Tracks memory operation metrics (latency, token usage, API costs) and provides analytics dashboards showing usage patterns, cost breakdown by provider, and performance trends. The system collects telemetry automatically without application instrumentation and exposes it through the Platform API and optional export to external analytics systems.

Automatically extracts entities and relationships from conversation text using LLM-powered NER/relation extraction, then stores them in graph databases (Neo4j, ArangoDB, etc.) to enable relationship-aware memory retrieval and reasoning. The system builds a knowledge graph where entities are nodes and relationships are edges, allowing queries like 'find all projects this user is working on' or 'what companies has this person mentioned'.

Provides two deployment models: a managed REST API platform (MemoryClient) for cloud-hosted deployments with built-in multi-tenancy and organizations, and an open-source self-hosted option (Memory class) for local deployments with full control over data and infrastructure. Both models expose identical memory operations (add, search, update, delete) through different client classes, allowing applications to switch deployment models with minimal code changes.

Provides AsyncMemory and AsyncMemoryClient classes that implement all memory operations (add, search, update, delete) as async/await coroutines, enabling concurrent memory operations without blocking. Built on Python's asyncio, the async implementation allows applications to perform multiple memory operations in parallel and integrate with async web frameworks (FastAPI, Quart, etc.) without thread pool overhead.

Abstracts LLM provider selection through an LlmFactory that supports 18+ providers (OpenAI, Anthropic, Ollama, Cohere, Groq, etc.), allowing runtime configuration of which model performs fact extraction, entity extraction, and other LLM-powered operations. Applications can specify provider and model name in configuration, and Mem0 handles provider-specific API calls, token counting, and response parsing without exposing provider details.

Enables memory queries filtered by session, user, agent, and custom metadata using a filter-based query system that applies constraints before semantic search. The system supports complex filter combinations (AND/OR logic) and allows retrieving memories scoped to specific conversation sessions or agents, preventing information leakage across isolation boundaries.

Allows updating existing memories with new information while maintaining audit trails and version history. The system tracks what changed, when it changed, and by which operation, enabling rollback capabilities and compliance auditing. Updates can modify memory content, metadata, or both, and the system handles re-embedding and re-indexing automatically.

Supports batch operations for adding, updating, or deleting multiple memories in a single API call, reducing latency and API overhead compared to individual operations. The system processes batches efficiently by grouping embeddings, database writes, and graph updates, and provides partial success semantics where some operations can fail without aborting the entire batch.

Allows applications to customize the LLM prompts used for fact extraction, entity extraction, and other memory operations through configuration, enabling domain-specific memory extraction tuned to application needs. Applications can provide custom system prompts, extraction instructions, and output format specifications that override defaults, allowing fine-grained control over what information is extracted and how it's structured.