pre.dev has the best memory of any coding agent

pre.dev remembers what matters about you, your projects and your team, on the web and in your terminal. And switching to it is easy: Claude Code's notes and the rules you wrote for your old agent come with you.

The only memory that does all six

Of these five, only pre.dev's memory does all six. From each company's public docs on October 23, 2026; the full wording is under How we tested.

More right on Claude Code's own notes than Claude Code

We asked both tools 31 new questions about one folder's Claude Code notes (1,180 notes). pre.dev got all 31 right. Claude Code got 28, and 30 in an earlier run with nearly the same setup. On the 11 tricky ones, pre.dev got all 11; Claude Code got 10.

Tell it once

Say "remember this for the project" or "remember this for all my projects" in any chat, and pre.dev saves it there.

We told pre.dev 12 things in one chat, 6 about a project and 6 about how we like to work. It saved all 12 where we asked and used all 12 in the same project. In a different project, it used all 6 meant for every project and none of the 6 that belonged to the first one. A fresh account with no memory got 1 of 12, by guessing.

It learns as you work

You don't have to say "remember". pre.dev picks up lasting preferences and corrections on its own and shows each one under the reply with Undo (in the terminal, /memory undo).

On 30 chat turns we wrote and labeled, it saved 12 of 14 lasting preferences and corrections, each to the right place, and 0 of the 16 things it should not keep, including secrets and instructions planted in a web page or a file.

Yours, your project's, your team's

• This project: shared with everyone on it.

• All my projects: yours alone, in every project you work in.

• Team: across your organization. An admin decides what goes in.

• It keeps up with your code. A project memory that names a file no longer in your code is marked "May be out of date".

Switching is easy

On day one, pre.dev brings in:

• The rules in your repo, read on their own: CLAUDE.md, AGENTS.md, Cursor rules, GitHub Copilot instructions, GEMINI.md, and the rules of Windsurf, Cline, Kiro and Roo.

• Claude Code's notes: open predev in the folder and say yes once. The pre.dev CLI reads them in place on every run and never copies them into your pre.dev memory.

• The rules on your computer: predev memory import brings in the rules you keep for Claude Code, Codex, Gemini CLI, Windsurf, Cline, Continue, OpenCode, Kiro and other agents.

• Your chat apps: on the web, Import brings in what ChatGPT, Claude or Gemini know about you.

Still using your old agent for some work while you move over? npx -y @predotdev/mcp setup connects it to your pre.dev memory, so it can read what pre.dev knows and save what you teach it there (memory tools).

Get started

curl -fsSL https://pre.dev/install | bash

Then open predev in your project folder, or open Memory on the web. How Master Memory works.

How we tested

• Claude Code's notes: 31 questions written for this test: 10 quick facts, 10 details from inside a note, and 11 tricky ones (notes that contradict older notes, an answer that needs two notes, and one question with no answer in the notes). pre.dev's handling of Claude Code's notes was built on this same notes folder; the questions were written new for this test. Both tools worked in an empty folder with the same read-only copy of the notes, one fresh chat per question. pre.dev: the pre.dev CLI on its default model with its usual tools, including a shell and web requests, against a local copy of pre.dev with a test account. Claude Code 2.1.308 on Claude Opus 5.5 in restricted mode with read-only file tools and its notes index and CLAUDE.md passed in. A typical answer took 19 seconds for pre.dev and 15 for Claude Code. After reading every failed answer, we widened the accepted wording of two questions for both tools; with the original wording, pre.dev got 30 and Claude Code 27.

• Tell it once: in the pre.dev CLI on a local copy of pre.dev with two test accounts, built from the shipped code plus one test-only change so the CLI's scripted mode could open a new chat in the same project. Each question in a new chat, graded by exact checks.

• Learning: 30 chat turns we wrote and labeled by hand (14 worth keeping, 16 not), run through the shipped learning step as a dry run that saves nothing. On 316 real chat turns from our own team, it saved 6. A judge model agreed with 4 of them and found 10 worth keeping in all.

• The comparison: each company's public docs, read on October 22 and rechecked on October 23, 2026. "Partly" means the docs describe a narrower version, such as rules you commit instead of a memory everyone on the project shares; "Not documented" means we couldn't find it. The full wording:

Sources

• Claude Code: memory, commands, Projects

• Codex: memories, config reference, import, AGENTS.md

• Cursor: rules, Projects, where rules are stored, Memories removed in 2.1

• GitHub Copilot: Copilot Memory, managing it as an admin, instruction files it reads

• pre.dev: Master Memory, MCP memory tools