Basic Memory and ModelBrain both keep an AI's memory on your own computer and both reach Claude, ChatGPT and Cursor over MCP. The real difference is who holds the pen. In Basic Memory, your assistant writes into the same Markdown files you write in. In ModelBrain, your assistant writes into its own store and only reads your notes. We call this the pen test: before you connect any memory tool to a folder you care about, find out whether the assistant can change the files in it.
Basic Memory stores memory as plain Markdown files, by default in a folder called ~/basic-memory, with a local SQLite index beside them. Its README says "AI and humans write to the same files; sync keeps them in step." Its MCP tools include ones to write, edit, move and delete notes, and it says Obsidian needs no setup because you can point Obsidian at the same folder. The result is one notebook that you and your assistant both edit.
ModelBrain keeps what your assistant saves in its own vault, separate from your notes. When you connect a folder or an Obsidian vault, ModelBrain reads it and indexes it so your assistant can search it, but it does not write to it. Our Obsidian support page lists "your vault files are never changed" as a tested item, including when you remove a vault. The result is two things: your notes, which only you edit, and a memory, which your assistant edits.
Neither design is wrong. They answer different needs.
If you want a notebook that your assistant keeps for you, shared writing is the point. You can open the file the assistant just wrote, fix a line, and the assistant sees your fix next time. Nothing is hidden in a format you cannot read.
If your notes folder is something you already care about, such as years of Obsidian notes, client files or a journal, shared writing has a cost. Any assistant you connect can reword, move or delete a note, and a mistake lands in your own files rather than in a separate memory you can throw away. Basic Memory guards against accidental overwrites and its cloud plan offers snapshots and point-in-time restore, so this is manageable. It is still a different promise from "the assistant cannot change this folder at all."
| Question | Basic Memory | ModelBrain 1.0.1 |
|---|---|---|
| Can the assistant change your notes? | Yes. Assistant and person write the same Markdown files | No. It writes to its own vault and only reads connected folders |
| Where memory is kept | Markdown files in ~/basic-memory plus a SQLite index, or its cloud | A vault folder on your computer |
| Can you read the memory without the tool? | Yes, it is plain Markdown | Through ModelBrain, or a full export at any time |
| Search | Full text plus vectors, optional reranking that is off by default | Hybrid search with a bundled reranker and learned ranker, on by default |
| Read-only access for one app | Not described on the pages we read | Yes, since 1.0.1: a hosted app can be allowed to search and read but not save or forget |
| Phone or browser with nothing to install | Yes, through its cloud | No |
| Source code | Open source, AGPL-3.0 | The engine is closed |
| Price | Free on your own machine. Cloud from $19 a month on its homepage | Free on your own computers. Sync between devices $6 a device a month billed yearly |
| Published retrieval benchmark | None on the pages we read | Yes, on the benchmarks page |
ModelBrain 1.0.1, released on November 18, 2026, added a way to let a hosted app in read only. Glama's health checker has been connected to our own ModelBrain account that way since that day. Its connection test passes, and the save and forget tools are not offered to it, so it can list and search but cannot add or remove anything. The pen test applies to apps as well as folders: you can decide which apps hold a pen at all.
Basic Memory is open source, so you can read and change all of it. Its cloud works from a phone or a browser with nothing to install, which ModelBrain does not offer. Its 1.0 release adds a team wiki, live notes, document import for PDF, Word, PowerPoint and CSV, and agents in beta. If you want your memory as files you can open in any editor, it gives you exactly that.
ModelBrain's engine is not open source today. Its memory is not a folder of Markdown you can open in any editor, so you read it through the app or an export. The free local vault is not encrypted at rest; the three readers a memory can have explains what that means. Syncing across computers is paid.
Make a copy of a small notes folder and connect only the copy. Ask your assistant to save a fact, then to correct it, then to forget it. After each step, check whether any file in the folder changed, using a diff or your file manager's modified date. If files changed, the assistant holds a pen in that folder. Then look for a read-only setting, and check whether it applies per app.
The test has a limit. It tells you who can write, not how well a tool finds things. For that, look for a published retrieval test, and for ours, see the benchmarks page.
For more tools side by side, see AI memory tools compared and seven AI memory tools compared by which assistants can reach the same memory.
Does Basic Memory change my Markdown files? Yes, by design. Its assistant tools write, edit, move and delete notes in the same files you edit, and its README says sync keeps the two in step.
Does ModelBrain write to my Obsidian vault? No. ModelBrain reads a connected vault and keeps what your assistant saves in its own vault. Its docs list "your vault files are never changed" as tested.
Is Basic Memory free? Running it on your own machine is free and open source under AGPL-3.0. Its cloud plan is paid; its homepage said from $19 a month on November 20, 2026.
Is ModelBrain free? The local product is free with no memory cap and no account. Syncing between your computers is $6 a device a month billed yearly, or $8 monthly.
Last checked November 20, 2026. Tools change often, so confirm a detail on the vendor's own page before relying on it.