Seven memory add-ons for Claude Code, Codex and Cursor, compared by what they store and where
A developer compared seven tools that give coding assistants a memory, and the real differences are where notes live and whether saving one costs a model call.
The short version
- The seven tools are claude-mem, Mem0, Cognee, Supermemory, Basic Memory, the official MCP memory server and hippo.
- Memories live in very different places: a local database, plain Markdown files, one JSONL file, or a hosted account.
- Only Mem0's hosted server requires an account; most others make accounts optional.
- The author builds hippo, so his own claims about it come with admitted limits.
- He found the tools by asking Claude and ChatGPT eight questions, which produced 16 answers.
Why it matters
If you use an AI coding assistant daily, you probably retype your project's rules and past mistakes at the start of each session. A memory tool can end that, but it also decides where your notes are kept and who can read them. This comparison gives you the questions to ask before you install one.
What was compared, and how
As dev.to reported, a developer set out to compare tools that give coding assistants a memory: notes that carry over from one session to the next, so you do not have to repeat your project's rules each time. He started by putting eight common developer questions to Claude and ChatGPT, such as which memory server is best for Claude Code. Those 16 answers named Basic Memory most often (9 times), followed by Mem0 (7), Cognee (6), Supermemory (6) and claude-mem (4).
He then read each project's own README and documentation on 22 November 2026 and linked to them so readers can verify his summaries. The seven tools in the table are claude-mem, Mem0, Cognee, Supermemory, Basic Memory, the official MCP memory server and his own project, hippo. MCP, short for Model Context Protocol, is a standard way for an assistant to plug into outside tools.
One caveat on method: the shortlist partly reflects what two chatbots happened to say, which is not the same as a survey of what developers use. The author also builds hippo, which came up only once in those answers, and he asks readers to keep that in mind.
Where your notes actually live
This is the biggest practical split. Basic Memory stores plain Markdown files on your own disk, which you and the AI can both read and edit. The official MCP memory server keeps a small knowledge graph (a web of linked facts) in a single local JSONL file, a simple text format with one record per line.
claude-mem uses a SQLite database (a small file-based database) in a folder in your home directory. The post notes that its Cursor install defaults to the project's hosted service, CMEM Pro, unless you choose another provider. Cognee runs locally by default, with a cloud version as an option.
Supermemory is the reverse: its Claude Code plugin talks to the hosted service by default, though a local mode keeps everything on your machine. Mem0's hosted server is the only one that requires an account. Its docs say nothing runs on your computer and the memories sit in your Mem0 account.
How a memory gets saved
The tools split into two camps. Some capture things automatically. claude-mem uses hooks, small triggers that fire as the assistant works, to pass each tool call to a second "observer" model that writes short notes and a summary when the session ends. Cognee and Supermemory also offer plugins that capture activity by themselves.
Others rely on the assistant deciding to save something. With Mem0 and the official MCP server, the model calls a function when it judges a fact worth keeping. Basic Memory works similarly through its note-writing tools, with automatic capture available only as an opt-in.
This matters because saving can cost a model call. claude-mem's observer and Supermemory's fact extraction both use a model, though each can run on a local one such as Ollama. Cognee runs local models with no key, and calls a provider only once you add a key. The post lists Basic Memory's extraction as not stated, with search embeddings computed locally by default.
Support for the three assistants also varies. Mem0, Supermemory and Basic Memory document setups for Claude Code, Codex and Cursor. The official server's README covers Claude Desktop and VS Code instead.
The author's own tool, with its limits
hippo keeps a SQLite store per repository plus a global one, and writes Markdown copies so rules can be searched, committed and reviewed like code. Its default save path makes no model call, and there is no account. Claude Code gets 9 hooks, Codex gets two, and Cursor gets 13 memory tools once the MCP setting is added.
The author reports a test he designed in advance across 29 repositories. A script corrected Claude Code once, and a few tasks later it had a chance to repeat the error. With hippo it repeated the error 11 times in 100, compared with 21 for Claude Code's built-in memory by itself, at the same cost.
He is open about the weak points. The test covered only Claude Code, so Codex and Cursor have no measured result. The improvement also showed up when hippo's lesson-saving was turned off, which means the test cannot show that the stored lessons caused it. And the test code and run records are not public. His project has 794 stars, against 119k for claude-mem. Treat the figures as the author's claim, not an independent benchmark.
How he suggests choosing
For mostly Claude Code and hands-off capture, he points to claude-mem and Cognee's plugin. If you would rather have someone else run the memory, Mem0's hosted server is the option. For notes you and your assistant both write and edit by hand, Basic Memory fits. The official server suits people who want a tiny local graph steered by their own prompts.
For one repository shared by two or three agents, he names Supermemory and hippo. The difference he draws is that Supermemory uses a model to extract facts and can be hosted or local, while hippo stays local without a model call on its normal save path.
The post's table also covers licences, and Basic Memory is the only AGPL-3.0 project on the list. It does not discuss prices, so check each project's site if cost matters to you.
What to watch
- Whether independent testers repeat the author's mistake-repeat comparison on Codex and Cursor, where he has no results.
- Whether the maintainers of the other tools correct anything in the post, since he invited fixes.
Try it
Before installing any memory tool, check three things in its docs: where the notes are stored, whether you can open and read them, and how you delete them.
Sources and links
- Reportingdev.to: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by dev.to · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
