Mem0
Listed at https://mem0.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Hobby 免费:10,000 memories + 1,000 retrieval API calls/月;Starter $19/月:50,000 memories + 5,000 calls | per million tokens |
| Output | Pro $249/月:Unlimited memories + 50,000 retrieval API calls + Graph Memory + 多项目;Enterprise 合同制 | per million tokens |
🤖 Supported Models (50)
✨ Pros
- ✓De-facto AI Agent memory layer (61,493 GitHub stars, Apache-2.0, mem0ai/mem0)
- ✓Clear 4-tier pricing: Hobby free → Starter $19 → Pro $249 → Enterprise
- ✓8+ LLM providers supported (OpenAI/Anthropic/Gemini/Mistral/Groq/Ollama/Bedrock/Azure OpenAI)
- ✓Graph Memory (Pro tier and up) for entity-relation modeling beyond pure vector retrieval
- ✓Complete SDKs: Python, TypeScript, Java, Go, Ruby — same API for self-host and Cloud
- ✓Local deployment supported: open-source + Docker — runs on China-hosted Kubernetes without proxy
⚠️ Cons
- ×Hobby tier only 1,000 retrieval calls/month — production typically needs Starter or Pro
- ×Pro tier $249/mo is expensive — same price point as Braintrust Pro but different positioning (memory vs evals)
- ×Graph Memory only on Pro and up — Hobby/Starter users get pure vector retrieval only
- ×No SSO/RBAC on free Hobby — Enterprise contract required for self-host with audit/SSO
- ×Retrieval quality depends on LLM-based fact extraction — ~200-500ms per memory vs pure vector search
- ×China-hosted Cloud SaaS access stability not yet verified (self-host bypasses entirely)
🎯 Best For
AI agent apps needing long-term memory (customer support bots, personalized assistants, multi-turn conversation systems); compliance-sensitive teams wanting self-host to avoid vendor lock-in; agent framework authors needing unified memory abstraction across multiple LLM providers
💰 Pricing & Plans
| Plan | Price | Memory / Scale | Best For |
|---|---|---|---|
| Hobby | Free | 10K add / 1K retrieval requests per month; 1 project; unlimited end users | Individuals exploring memory-augmented side projects |
| Starter | $19 / month | 50K add / 5K retrieval requests per month; 1 project; community support | Solo builders shipping their first AI agent |
| Pro | $249 / month | 500K add / 50K retrieval per month; unlimited projects; Graph Memory + Dream consolidation; advanced analytics; private Slack | Production apps with real scale and multi-project needs |
| Enterprise | Custom | Unlimited add & retrieval requests; unlimited projects | Teams needing compliance, control & SLAs |
| Enterprise add-ons | Included in contract | On-prem deployment; audit logs; custom integrations & SSO; SLA support | Regulated / self-hosted enterprise stacks |
| Usage-based | Custom (metered) | For traffic that doesn't map cleanly to a fixed tier | Bursty or unpredictable workloads |
| Open Source (self-host) | $0 (Apache-2.0) | Self-managed; Docker Compose; same API surface as Platform | Teams avoiding vendor lock-in; China-hosted deployment |
🔧 API & Developer Experience
- •SDKs: Python (`pip install mem0ai`) and Node (`npm install mem0ai`) — same API for Platform and Open Source
- •CLI: `pip install mem0-cli` or `npm install -g @mem0/cli`; `mem0 init --agent` mints an evaluation API key in under 5 seconds, no email/OTP
- •Adoption & license: 62,590 GitHub stars, Apache-2.0 (mem0ai/mem0) — de-facto standard for agent memory
- •LLM flexibility: Platform handles providers server-side; 8+ supported (OpenAI, Anthropic, Gemini, Mistral, Groq, Ollama, AWS Bedrock, Azure OpenAI)
- •MCP & integrations: Live MCP server at mcp.mem0.ai; OpenMemory Gateway, CLI, REST API and published OpenAPI spec
- •Memory pipeline: ADD-only extraction → storage across vector / entity / history stores → multi-signal retrieval; async client for concurrent calls
- •Advanced features: Graph Memory (entity linking), Dream memory consolidation, multimodal input (images/PDFs), custom categories, webhooks, feedback mechanism
- •Billing & trial: Free Hobby tier; paid Starter/Pro monthly; Enterprise custom; usage-based metering for irregular traffic; self-host is free
🧠 Memory Layer for AI Agents
Mem0 is a persistent memory layer that gives LLM agents context that survives across sessions, so a chatbot remembers a user's preferences, past tickets and running projects instead of starting from zero every call. Its pipeline extracts durable facts from a conversation (ADD-only distillation), writes them into vector, entity and graph stores, and returns a compact, relevance-ranked memory block that is injected into the prompt — cutting token spend while raising answer quality versus stuffing raw history. Graph Memory links facts through shared entities for entity-centric, multi-hop questions, and Dream consolidates overlapping memories over time. The same SDK drives both the managed Platform and the Apache-2.0 Open Source build, so teams can prototype on Cloud and later self-host without rewriting a line of integration code.
🌐 China Access & Latency
Mem0's managed Platform (api.mem0.ai) is served from global cloud regions with no mainland China node, so calling it directly from China carries cross-border latency and intermittent connectivity risk — treat Cloud as the fast path for prototyping from outside or via an overseas relay. The decisive advantage for China-based teams is the Apache-2.0 Open Source build: it ships as a self-hosted service (Docker Compose with auth, API keys and a setup wizard) that runs on mainland Kubernetes or a domestic VM with zero outbound dependencies. When self-hosted, embeddings and the extraction LLM can both point at domestic providers or a local Ollama instance, keeping every memory read and write entirely onshore — which also resolves the data-residency and compliance concerns that make a foreign memory SaaS hard to adopt for regulated workloads.