Guide · Agent memory

What each memory design lets you take with you

Four ways agents store what they know about you, against the three questions that decide whether you can leave. Hover any cell.

Design
Bulk export
Read by another tool
Inspectable by you
Built-in assistant memoryA silent profile the product maintains
~
Sometimes an export exists, sometimes only a summary — and it is rarely complete enough to reconstruct the detail.
—
The format belongs to one product. A second agent cannot be pointed at it, so switching means starting over.
—
You see whatever the product chooses to show. There is no file to open and no diff to inspect.
Per-project notes fileA Markdown file beside the work
✓
It is your file. Copy, zip, email or commit it — no export step exists because none is needed.
✓
Any agent that can read Markdown can use it. This is the design that survives switching tools.
✓
Open it in any editor, read it line by line, and fix a wrong belief by editing the sentence.
Self-hosted memory storeA database or vector store you run
✓
You control the store, so a full dump is always possible — though it takes a script rather than a button.
✓
Point a second agent at the same store over MCP and it inherits everything the first one learned.
~
Queryable and inspectable, but it is data with a schema — not prose you read at breakfast.
Managed memory over MCPA hosted service your agent calls
~
Depends entirely on the vendor. Some expose full records, others only what their API chooses to surface.
✓
Portable in principle — the protocol is the interface, so another MCP client can usually attach.
—
You see what the service shows. No files, no diffs, and no way to audit a belief you never asked about.
✓Satisfied ~Depends on the vendor or takes real work —Not realistically available

Assessment based on the documented export behaviour of each design, not a benchmark. The two columns that decide portability are the two where storage is yours: a file you can open, or a store you run.