What Claude's built-in memory actually does
On claude.ai, Claude maintains a memory summary drawn from your conversations: you can read, edit, and delete it in your Claude settings, and memory is scoped so that separate projects keep separate context. Files you add to a Project are available in every chat inside that Project. Claude Code, the terminal agent, reads Markdown memory files (CLAUDE.md and auto-memory notes) at the start of each session. All of this is real, documented behavior: Anthropic's own docs describe each layer, and within its scope each one works.
Importing your memories from ChatGPT
Coming from ChatGPT, you don't have to start from zero (what ChatGPT's own memory does is covered in the ChatGPT memory guide). Claude has an import for this, available on every plan on the web and in Claude Desktop, according to Anthropic's help page. You have ChatGPT write out what it has stored about you, using the prompt on that page, and paste the text into Claude under Settings → Your Claude → Start import. Claude turns it into individual memory entries you can review afterwards. Anthropic still calls the feature experimental. Only the stored facts move over, not your chat history, and Claude mainly keeps what relates to work. Your documents don't come along.
Where it stops: three limits
- Project boundaries. The memory summary lives inside claude.ai and its project scopes. What Claude learned in one project doesn't follow you to the next, and none of it reaches Claude Code, ChatGPT, or Cursor.
- No documents. Chat memory summarizes conversations. The contract PDF, the scanned invoice, the mail thread with the decision in it: none of that is in the summary, and Project uploads stay locked to their Project.
- No citations. A summary compresses. Ask Claude where a remembered fact came from and there's nothing to point to. The original context is gone.
Claude Code: CLAUDE.md and auto memory
Claude Code has two layers of its own, both Markdown files on your machine. CLAUDE.md is the one you write: instructions for every project in ~/.claude/CLAUDE.md, for one project in ./CLAUDE.md (shared with your team through Git), or just for you in ./CLAUDE.local.md. You pull in further files with @path. Auto memory is the one Claude writes, on by default: corrections, preferences and project state go into ~/.claude/projects/<project>/memory/, one directory per Git repository, indexed by MEMORY.md. Every session reads the first 200 lines of that index. You can switch it off with /memory. Both stay with your machine and the repository: what Claude Code learns there, claude.ai never sees, and the other way round. The details are in Anthropic's docs.
The DIY route: CLAUDE.md, Markdown files, Git
The most popular answer to these limits is to build your own: keep your knowledge in Markdown files, point Claude Code at the folder, version it with Git. It's a good answer: free, local, transparent, and for plain text it genuinely works. Half the Claude Code community runs on some variant of it, from a single CLAUDE.md to elaborate memory directories. If your knowledge is text you wrote yourself, you can stop reading here.
Where the DIY route breaks: scanned documents
Real knowledge doesn't arrive as tidy Markdown. It arrives as PDFs, receipts, contracts, scans, mail attachments. Markdown folders have no answer for those: a scan in a folder is a file name, not knowledge. Users of these systems say it themselves; in a Mac Power Users thread, an Evernote emigrant described Obsidian as “good at text only/minimal images”, with getting receipts and scans in being “clunky”. That's the honest boundary of the DIY route: it holds what you typed, not what you received.
claude-mem and other plugins compared
Between building your own and a hosted memory sit plugins. The best known is claude-mem (open source, Apache 2.0). It hooks into Claude Code and other coding agents, records what the agent does in a session, has an AI summarize it, and hands the next session the relevant context. Storage is local, in SQLite with vector search; a cloud backup is optional. For programming work that's strong: tomorrow the agent knows what it did in the code yesterday. It isn't built for your paperwork. It remembers sessions, not contracts or scans, and it works in the terminal, not in claude.ai or ChatGPT. Similar plugins and local memory servers mostly follow the same pattern.
| Where it runs | What it remembers | Documents with citations | |
|---|---|---|---|
| Memory in claude.ai | claude.ai and Claude Desktop | A summary of your conversations, separate per project | No |
| Claude Code (CLAUDE.md, auto memory) | Local, per repository | Your instructions and what Claude learns while working | No |
| claude-mem | Local, in coding agents | Summaries of past sessions | No |
| Memexa | Hosted in the EU, in any AI app that speaks MCP | Notes, documents and scans as a linked knowledge graph | Yes, down to the page |
What a document memory looks like
A memory that includes documents has to read them. Memexa runs every upload through OCR, splits it, embeds it, and distills it into linked notes in a knowledge graph, so a search hit can say “contract, page 4” instead of “somewhere in your files”. Claude reads and writes this memory over MCP, and so does every other AI app that speaks MCP. Connecting takes one URL:
Claude Code
claude mcp add --transport http --scope user memexa https://brain.memexa.ai/mcp
Claude Desktop / claude.ai
Settings → Connectors → Add custom connector → https://brain.memexa.ai/mcp
This works on every Claude plan, the free one included: it allows exactly one custom connector, Pro and Max allow several. Open the dialog in Claude. Sign in once in the browser (OAuth 2.1), no tokens to copy. To have ChatGPT and Claude read the same memory, see Give ChatGPT and Claude the same memory; Mistral Vibe and Cursor are on the how-it-works page. The full walkthrough for claude.ai, Claude Desktop and Claude Code is in Set up Memexa in Claude; if you use Obsidian, see three ways to connect Obsidian and Claude. Keep the built-in memory and the Markdown folder. They don't conflict. The document layer is what they're missing.
Where your data lives
A memory is a strange thing to rent, so the exit has to be real: Memexa hosts in the EU under the GDPR, your content is never used for training, and everything exports with one click, notes as Markdown. The graph is an index, not a lock-in. Details in the privacy policy.