AI Agent Memory vs Shared Notes: What Each Agent Remembers and What It Reads (2026)

Agent memory is what ChatGPT, Claude, Gemini, Claude Code, Codex and Copilot decide to keep about you, and each one keeps it inside its own product, account or machine. Shared notes are what you write once so every agent and every teammate reads the same current facts. Use memory for habits, and notes for your business.

I run my work across ChatGPT, Claude, Claude Code and Codex every day, so I opened each vendor's own memory and instruction pages on 26 October 2026 to see what each agent keeps, where it lives and who else can read it. The sources are listed at the end. Here is what they say today, and where shared notes pick up.

What each agent remembers, at a glance

Checked on 26 October 2026 on each vendor's own help pages and docs. "You write" means text you control; "it keeps" means what the agent saves by itself.

The pattern holds in every row. Each feature makes one product better at knowing you or your repository. None of them is a place where your other agents, or your other people, read the same business facts.

Memory is what the agent decides to keep

The first thing the pages agree on is who holds the pen.

• ChatGPT: OpenAI says Memory "does not retain every detail from every conversation" and that ChatGPT decides which available information is relevant to a response. The memory summary is a high-level view and may omit details.
• Claude: memory is saved as individual topics while you chat. Claude can save on its own, or you can ask it to remember something.
• Claude Code: auto memory is notes Claude writes itself. Anthropic's docs say it doesn't save something every session and decides what is worth remembering. It skips anything it can derive from the code or that your CLAUDE.md already says.
• Codex: it creates local memories from eligible past chats in the background, after a chat has been idle, and redacts secrets from generated fields. OpenAI's page tells you to keep required team guidance in AGENTS.md or checked-in docs, because memory is "a helpful recall layer".
• GitHub Copilot: repository facts come from actions by people with write access who have Memory enabled. Entries that go unused for 28 days are deleted automatically.

That is the right design for habits: your tone, your preferred tools, the way you like a code review. It is the wrong place for a fact that must be true every time, such as a price, a refund rule or the words you never use. An agent can decide a fact isn't relevant, or never see it.

Where each memory stops

Every limit below comes from the vendor's own page.

It stays on one machine. Claude Code's docs say auto memory is machine-local: files are not shared across machines or cloud environments. Codex's local memories live in a store under ~/.codex and are separate from ChatGPT web memory.

It belongs to one person. Claude memory is per user, and owners can't view or edit individual members' memories. In ChatGPT, shared projects automatically use project-only memory and can't see a member's memories from outside the project. That protects privacy, and it also means a teammate's agent never inherits what yours learned.

It belongs to one vendor. Claude can import memory from other assistants and export its own, and Anthropic labels the feature experimental. It is a one-time copy: it doesn't keep two memories in step.

It isn't a document you can audit. OpenAI's memory summary doesn't necessarily include every detail or source memory can use. Deleting a ChatGPT chat doesn't necessarily delete a saved memory created from it, and Anthropic says deleting a conversation doesn't remove the memory entries created from it.

It depends on your account type. Gemini's memory needs a personal Google Account, so work and school accounts are out. Claude Team and Enterprise members' memory stays off until an owner enables it and the member turns it on.

None of this makes memory a bad feature. Keep it on for personal preferences. The gap is that "what my business knows" isn't a personal preference.

Instruction files are the middle layer

Between memory and shared notes sit the files you write yourself, which the coding agents read at the start of a session.

• Claude Code reads CLAUDE.md at user, project and organization level, and its docs now say it can read a repository's AGENTS.md when there is no CLAUDE.md in your working directory or above it. That needs Claude Code v2.1.277 or later. The docs target under 200 lines per CLAUDE.md.
• Codex reads AGENTS.md from ~/.codex and from the project root down to your working directory, one file per directory. It stops adding files once the combined content reaches 32 KiB by default.
• Cursor supports project rules in .cursor/rules, user rules, team rules on Team and Enterprise plans, and AGENTS.md, including nested files.

These are written by you, versioned in git and shared with whoever clones the repository. They are the right home for build commands and code rules. They live in a repository, so a marketing agent that never opens your repo never reads them, and your positioning and prices don't belong in a code repository anyway. I covered where each coding agent looks in AGENTS.md vs CLAUDE.md.

What shared notes add

Shared notes are the third layer: facts you write once, kept in one place every agent can reach. In LotsNotes that means:

• You write them, or your agent saves them when you ask. LotsNotes doesn't write notes on its own. A note changes when you edit it or when an agent saves it on your instruction, and every save keeps a version you can restore.
• Any agent reads the same note. Claude, ChatGPT, Codex, Cursor, cloud agents and your own code connect over MCP or REST. Workspace notes come back with every folder's notes, so a workspace note titled AGENTS.md reaches every agent, and a folder can hold its own.
• People and agents share roles. Owners, admins, members and viewers work within the same permissions. Private notes stay on lotsnotes.com and are never returned to agents or the API.
• You can see and cut off access. The Connected apps list shows which agents have access and when each last used it, and you can disconnect one everywhere.

Our own AI agents keep their instructions, plans and handoffs in LotsNotes, and Claude, Codex and cloud agents read them from there.

When I called find_context through the LotsNotes MCP on 26 October 2026, the result came back in two lists, workspace_notes for the notes that apply everywhere and folder_notes for the folder I asked about, each with its lotsnotes.com link. That is the whole read path: an agent asks by title, gets the current version and reads only what it needs.

A prompt that works in any connected agent:

Use LotsNotes. Find the context for the "Acme" workspace, "Marketing" folder. Read the AGENTS.md notes first, then the brand voice note. Draft three launch emails in that voice. If you learn something new about our customers, tell me what you'd add to the customer insights note and wait for my yes.

Behind it the agent calls lotsnotes_find_context and lotsnotes_read_note, and lotsnotes_save_context_note only when you say yes. The save takes an expected_version, so if someone changed the note since the agent read it, the save fails with a conflict instead of overwriting, and the agent reads again and merges.

Put each thing where it belongs

One line at the top of your personal instructions connects the layers: "For anything about my business, read the AGENTS.md note in LotsNotes first." If you work with more than one client, the client folders guide shows how to keep each brand apart.

Set it up in five minutes

1. Create a workspace in LotsNotes and add one note titled AGENTS.md with your positioning, customers, voice and rules.
2. Connect your agent from the setup page. In Claude Code: claude mcp add --transport http lots-notes https://api.lotsnotes.com/mcp, then run /mcp and choose Authenticate. In Codex: codex mcp add lots-notes --url https://api.lotsnotes.com/mcp, then codex mcp login lots-notes. Claude, ChatGPT and Cursor have their own steps on the same page.
3. Add the one-line pointer to your agent's instructions, and keep its memory on.

LotsNotes is free for one person: one workspace, three folders, unlimited notes and checklists, and MCP and REST access. Pro adds unlimited workspaces, folders and team members with roles, and it is included while you have a Pro or Business plan of LotsSocial, LotsBlog, LotsTeam, LotsCRM, LotsInvoice, LotsBusiness or LotsStore. Each agent's built-in memory comes with its own plan at no extra cost, so nothing here asks you to give any of it up. Prices are on the pricing page.

FAQ

What is AI agent memory?

It is information an AI agent saves from your work and reuses later, such as preferences, past chats or notes it wrote itself. ChatGPT, Claude, Gemini, Claude Code, Codex and Copilot each have their own version. Each is kept inside that product, account, machine or repository.

Is agent memory the same as a shared notes workspace?

No. Memory is chosen and written by the agent, and it is personal to you in most products. A shared notes workspace holds text you or your team wrote, with versions and roles, and any connected agent can read it.

Can Codex or Cursor read my ChatGPT or Claude memory?

Not directly. OpenAI says Codex's local memories are a separate store from ChatGPT web memory, and Claude Code's auto memory stays on one machine. To give every agent the same facts, keep them in a place all of them can read, such as LotsNotes over MCP.

Does Claude Code read AGENTS.md?

Yes, in current versions. Anthropic's docs say Claude Code reads a repository's AGENTS.md when you have no CLAUDE.md in your working directory or above it, from v2.1.277. If you have both, it reads your CLAUDE.md files unless you change the setting.

Does LotsNotes remember things on its own?

No. LotsNotes stores the notes you write or tell your agent to save, and it keeps a version of every save. The agent decides nothing about what is kept unless you ask it to save.

Can my agents read my private notes in LotsNotes?

No. Private notes stay on lotsnotes.com and are never returned through MCP, the API or any sync. Agents see only shared notes, within your workspace roles.

Sources

Opened on 26 October 2026:

• OpenAI, Memory in ChatGPT
• OpenAI, Projects in ChatGPT
• OpenAI, Memories (ChatGPT and Codex)
• OpenAI, Codex AGENTS.md
• Anthropic, Use Claude's chat search and memory
• Anthropic, How Claude remembers your project
• Google, Get personalization with memory of your past Gemini chats
• Cursor, Rules
• GitHub, About Copilot Memory
• LotsNotes, pricing, llms.txt and API docs

Write it once

Let each agent remember your habits. Put the business context every agent needs where every agent can read it: connect your agent to LotsNotes and start with one AGENTS.md note. For other places to keep it, see the best MCP memory servers and the Claude Projects alternatives.