Source: raw/3_Levels_of_Claude_MEMORY_Clearly_Explained_how_to_setup.md (primary). Earlier state of the memory feature from the same creator: raw/How_to_Transplant_Your_ChatGPT_Brain_into_Claude_simple_updated_version.md (2026-03-04, memory import) and raw/Claude_vs_ChatGPT_-_Which_is_ACTUALLY_better_my_updated_2026_review.md (2026-04-13, memory format).

Creator: Eliot Prince URL: https://www.youtube.com/watch?v=EV-NSTvb1kQ Duration: 16:23 Platform: YouTube Published: 2026-08-31

Eliot Prince walks through the late-August 2026 redesign of the Claude app’s memory, which added a Memory tab, a structured format and a deadline to export the old memory. He then sets out three levels of memory. Level 1 is account memory, which you should build on purpose instead of letting it pile up. Level 2 is Projects, with their own context files and a project memory that regenerates daily. Level 3 is a memory.md file in a local folder that Claude reads and updates each session, and that any tool can use. This is the Claude.ai and Cowork counterpart to Claude Code’s file-based Auto Memory.

Key Takeaways

  • Memory moved to its own settings tab and changed format. Settings → Memory now splits what Claude knows into sections: You (profile), Topics, Areas, People and Interests. The old memory was “a long big blob of text.” It now applies across both chat and Cowork; before, it worked only in chat.
  • Export the legacy memory before the cutoff. The new tab showed a migration notice with a countdown (10 days at recording, so around 9–10 September 2026 by Eliot’s estimate). It offered an export as a Markdown file he calls “Claude legacy memory.md.” That window has probably closed by now.
  • Audit what got saved. His memory listed “likes oranges,” which was untrue and probably saved from a throwaway demo. Memory left to accumulate gets stale, mixes up multiple businesses, and keeps random facts. He prefers to delete it and rebuild it on purpose.
  • Account memory is shallow by design. It holds who you are and what you work on. It doesn’t hold the state of each client or project, which is what Level 2 is for.
  • Project memory regenerates every 24 hours from that project’s chats. It is private to you even on a team account. It records current state, what is on the horizon, and key learnings and principles / approach and patterns. Because it updates overnight, today’s decisions won’t show until tomorrow.
  • A memory.md file in a folder works across tools. Plain files in your own folder can be read by Claude Code or another vendor’s assistant (he names ChatGPT, Gemini and Grok). Account and project memory stay inside Claude.

Level 1 — Account memory, built deliberately

  1. Open Settings → Memory. If the tab isn’t there, it may still be rolling out to you.
  2. Read every section and use the editor at the bottom to correct or delete entries. Expect some stale or random facts.
  3. To rebuild it, Eliot runs a 10-question interview prompt from his recipe vault (see AI Recipe Vault; the prompt text isn’t in the transcript). It starts by asking for your CV or a LinkedIn profile exported as PDF. He picks LinkedIn because it is already public, so nothing sensitive goes into memory. Then come follow-ups: what you do in your customers’ words rather than your job title, your goals, what you love and hate doing, and what you use AI for. He answers by voice dictation.
  4. The interview produces a memory doc (background, targets, people, tool stack). When it reads right, tell Claude to “commit this memory doc to memory,” and it updates the structured sections.

Coming from ChatGPT. Claude has a built-in import, documented by Eliot in March 2026: Settings → Capabilities → “Import memory from other AI providers” → Start import. Claude gives you a prompt to run in ChatGPT; you paste the code block ChatGPT returns back into Claude and click “Add to memory.” Full walkthrough: Moving from ChatGPT to Claude. Audit the result the same way, since an import brings in whatever ChatGPT had saved.

How the format changed. In April 2026, Eliot described Claude’s memory as “one long flowing document,” unlike ChatGPT’s short separate saved memories. At the time he preferred it because it held more context. By the August redesign he was calling that same blob of text “not optimal for AI to read.” The new sectioned layout is closer to ChatGPT’s approach.

Level 2 — Projects as sandboxed memory

  • A Project only uses its own material. It answers from its files, chats, instructions and memory (plus web search if you allow it), so one client’s details don’t leak into another’s work.
  • What goes in a project’s context. For a consulting client, Eliot keeps an executive summary and business profile, a “business DNA” file about himself, his consulting frameworks and playbooks, and a transcript of every client call.
  • Read the project memory before each session. The current-state section (for example: proposal delivered, project configured, discovery call prepped) is the quickest way back into a long-running engagement.

See Projects for the feature itself and Skills vs. Projects for when a repeated process should be a skill instead.

Level 3 — A memory.md second brain

Eliot’s setup, recorded before the September merge:

  1. Create a Cowork project tied to a new local folder (his example: second-brain-os).
  2. Run a prompt that creates memory.md at the folder root. He describes it as “the living context file for this work,” with these sections and rules:
# memory.md
## Purpose
## Current state
## Decisions and why
## What's next
## Key links / reference tools
 
<!-- Rules: read this file at the start of every session.
     Update it after any big decision, at the end of a session,
     and when a major deliverable is complete. -->

The block above is reconstructed from his description of the prompt, not copied from it.

  1. Nest it. In his full system, each project subfolder has its own memory.md with that project’s current state, and a top-level memory.md holds the overall context. Pointing a tool at one subfolder gives it that project’s memory plus its knowledge base.
  2. Let it update as you work, or ask for an update. For example: “update memory.md: this is my main second brain.”
  • Since the merge. Chat and Cowork merged on 2026-09-16 and there is no longer a mode to pick. The “select Cowork mode” step now just means creating a project and attaching a local folder (see Claude Cowork).
  • Same pattern as Claude Code. This memory.md does the same job as a project’s CLAUDE.md plus Auto Memory’s MEMORY.md index in Claude Code. The difference is that it is one file you control, organised around decisions and next steps. See Claude Code Memory Architectures Compared.

Try It

  1. Export, then audit. Download whatever export the Memory tab still offers. Then go through You, Topics, Areas and People and delete anything wrong or out of date.
  2. Write a one-page memory doc yourself (role, customers in their own words, goals, tools, key people) and ask Claude to commit it. Ask “what do you remember about me?” a day later to check it took.
  3. Move one client into a Project with an executive summary, your frameworks and the last three call transcripts. Read its project memory the next morning.
  4. Start a memory.md in one working folder with the five sections above. Begin each session with “read memory.md,” end it with “update memory.md,” and do this for a week.

Open Questions

  • First-party documentation. The Memory-tab redesign, the export deadline and the 24-hour project-memory cycle all come from one creator’s screen recording. This wiki has no Anthropic announcement for the August 2026 memory migration.
  • Plan availability. He doesn’t say which plans get project memory or the new Memory tab. Projects records memory as free and past-chat search as paid, which predates this redesign.
  • The interview prompt itself is behind his recipe-vault link and isn’t captured here.
  • Did memory import survive the redesign and the September merge? The March 2026 path ran through Settings → Capabilities. The August video shows memory in its own Settings → Memory tab and doesn’t mention an import. Where the import lives now hasn’t been confirmed.