Claude Code auto memory
- Where it lives and what it covers
- Markdown files for each git repository, on that machine.
- How it finds things
- Loads the first 200 lines or 25 KB of MEMORY.md, then reads other files with its standard file tools.
Built-in AI memory, compared
Claude Code memory vs CortexMind
Claude Code, Codex, claude.ai and ChatGPT already remember some things. Here is where each memory lives, what it saves, and how Claude Code's memory and CortexMind compared in our side-by-side tests.
39%Claude Code auto memory
89%CortexMind
Questions each could still answer in a new conversation, after the same working sessions.
Checked 9 November 2026. These products change, so follow the sources.
What each one is
Codex memories
- Where it lives and what it covers
- Off by default. Stored locally on each machine, separate from ChatGPT memory.
- How it finds things
- Not stated in the page cited.
claude.ai memory
- Where it lives and what it covers
- A separate system on web, desktop and mobile. You can view and edit it.
- How it finds things
- A separate search of past chats uses retrieval (RAG).
ChatGPT memory
- Where it lives and what it covers
- Saved memories and chat history reference, varying by plan and region.
- How it finds things
- Not stated in the page cited.
CortexMind
- Where it lives and what it covers
- One database file on your computer, used by every tool you connect, across all your projects.
- How it finds things
- Keyword and meaning search together, then a second model reranks the best candidates.
The biggest difference is what gets saved
Claude Code's documentation says it "doesn't save something every session" and decides what is worth remembering. It also says it skips what it can derive from your code, including debugging fixes. CortexMind's instructions tell your AI to save every exchange in full, and to pull out lasting facts alongside it.
So we tested it. After working sessions on an empty memory, we asked what each one could still answer in a new conversation.
Claude Code auto memory CortexMind
When you say remember, both remember. The gap is everything you did not think to flag at the time, and no search can find what was never saved.
Even if both had saved the same things
Most of the gap is decided at saving. But suppose built-in memory had kept everything CortexMind keeps. We tested that too, with the same notes in both.
Claude Code's auto memory loads the start of an index, then reaches everything else with its standard file tools. CortexMind searches its whole memory by keyword and by meaning. On a very small memory there is little to choose between them. As a memory builds up through serious, long-term use, less and less of it loads on its own, and CortexMind pulls ahead.
The first bar shows what that means in practice: Claude Code working only from what it normally saves.
Claude Code, with what it normally saves Claude Code, if it had saved everything CortexMind
Very small
Search about the same, far less to find
Months of serious use
CortexMind ahead either way
A very large memory
CortexMind further ahead
Where CortexMind goes further
It can keep growing
CortexMind sets no limit on how many memories it keeps, and every search looks across all of them.
One memory for every project
Claude Code's memory stays with one repository, so a question asked from a different project finds nothing there. In our tests, CortexMind found the answer every time.
One memory for every tool
Claude Code, Codex, Cursor, ChatGPT and claude.ai can all connect to the same memory. None of the built-in memories above documents sharing with another company's tools.
You can see it and correct it
A dashboard on your computer shows every memory. Change what matters, and when a fact goes stale, replace it so the old one stops coming back.
About these tests
We ran hundreds of side-by-side test sessions in Claude Code, on memories of up to 23,000 notes. November 2026. These products change and can vary by account, so your results may differ.
