Compare
total-agent-memory vs Mem0
Mem0 is a memory layer you build into your own app and needs an LLM to run. total-agent-memory is an MCP server for Claude Code, Codex CLI and Cursor that runs on your machine without one. What each is for, what the benchmarks do and do not show.
Facts about Mem0 were checked on October 9, 2026 against its public source. Facts about total-agent-memory are current for the version in the site header.
The short version
Mem0 is a memory layer for applications: you call its SDK from your own code, and it extracts, stores and retrieves memories for the users of your app. It needs a language model to work, with OpenAI’s gpt-5-mini and text-embedding-3-small as the defaults, and it comes as a Python or Node library, a self-hosted server, or a hosted platform.
total-agent-memory (TAM) is a memory server for coding agents you already use. Claude Code, Codex CLI, Cursor and other MCP clients call it through the Model Context Protocol. It keeps everything in one SQLite file on your machine, needs no API key, and makes no LLM call on the save or search path.
If you are building a chatbot or a product with many users, Mem0 is the tool. If you want your own coding agent to remember your decisions across sessions, TAM is.
Side by side
| Mem0 | total-agent-memory | |
|---|---|---|
| What it is | SDK and platform for apps you build | MCP server for coding agents |
| How you use it | pip install mem0ai, then call it from your code | One command per client, then talk to your agent |
| Needs an LLM to run | Yes; gpt-5-mini by default, other providers supported | No; optional, off the hot path |
| Embeddings | OpenAI text-embedding-3-small by default | Local MiniLM model, downloaded once |
| Where data lives | Your vector store, the self-hosted server or Mem0’s cloud | One SQLite file in ~/.tam/ |
| Team use | Self-hosted server or hosted platform | Team server with personal, team and shared scopes |
| Licence | Apache-2.0 (library); platform is a paid service | MIT, everything included |
| Community | 66,891 GitHub stars | 72 GitHub stars |
Facts about Mem0 come from its README on the date above. If something is out of date, open an issue and this page gets corrected.
What the benchmarks say
Mem0 publishes 92.5 on LoCoMo and 94.4 on LongMemEval for its new algorithm. Those numbers describe the managed platform with its production model stack and an LLM judge; they are end-to-end answer accuracy.
TAM publishes two kinds of numbers. On LongMemEval we report retrieval recall: R@5 of 95.1% on the 470 non-abstention questions, measured on version 13 through the real search path, with the per-question file public. A retrieval number and an answer-accuracy number cannot be compared, in either direction, so we do not.
The one like-for-like run we have is LoCoMo QA with the same prompt and the same judge for both systems, done during development of 14.0: TAM 72.21%, Mem0 OSS 71.36%, paired difference +0.84 points with a 95% interval of −3.03 to +4.96. No quality advantage is established, and we say so on the benchmarks page. Mem0 OSS in that run used a local Qdrant and FastEmbed setup, not the hosted platform.
When to pick Mem0
- You are building a product and want memory for its users, with SDKs, a dashboard and a managed option.
- You already pay for an LLM provider and are fine with memory extraction going through it.
- You need the hosted platform’s features and support.
When to pick total-agent-memory
- You use Claude Code, Codex CLI, Cursor or another MCP client and want it to remember decisions, fixes and conventions between sessions.
- Your code and notes must stay on your machine or your own server; no key, no cloud, no model in the loop.
- You want the benchmark files, not only the headline number.
Install takes one command: see the quick start.