Tavimi

Guides

Why Claude (and ChatGPT) forgets everything between conversations, and how to fix it

An AI assistant only works with the text of the current conversation, its “context window”, so a new conversation starts empty. Claude and ChatGPT now offer a memory that carries over part of your past chats, but it belongs to each user and stays incomplete. To stop re-explaining everything, write your context down once (instructions, a project, a file or an MCP server) and, as a team, keep it in a shared space that every person and every assistant can read.

Published

Yesterday you spent an hour with Claude or ChatGPT preparing a quote for a client: their background, their constraints, the tone to use, the two options you ruled out. This morning you open a new conversation and the assistant knows nothing. A colleague opens the quote with their own AI assistant: it analyzes the document, but suggests a different version, because it doesn’t know the context. Nothing is broken. This is how these tools work, and there are several simple ways to stop explaining everything again.

What’s going on: the context window

A language model doesn’t “keep” anything between two replies. With every message, it rereads all the text it is given: your messages, its own replies, attached files and any instructions. Anthropic calls this the context window and describes it as the model’s “working memory”, separate from the data it was trained on. As a conversation goes on, each exchange is added to that window.

This has three practical consequences:

  • A new conversation starts with an empty window. What you said yesterday isn’t in it, unless some mechanism puts it back (memory, a project, instructions).
  • One very long conversation is not the answer. The window has a size limit. Anthropic’s documentation explains that interfaces like claude.ai may manage the window on a “first in, first out” basis, and that accuracy drops as the context grows.
  • What matters is what goes into the window. Short, accurate, up-to-date context beats a long history.

Every fix below does the same thing: it automatically puts the right text back into the window at the start of each conversation.

ChatGPT memory and Claude memory today

Both vendors have added a memory that links conversations together. They work on the same principle, with a few differences.

ChatGPT

OpenAI’s help center describes two mechanisms, both set in Settings > Personalization:

  • Saved memories: details you explicitly ask it to remember, for example “remember that I work at an agency and write for industrial clients.”
  • Reference chat history: ChatGPT draws on your past conversations to pick up useful information.

A memory summary shows what ChatGPT remembers, but OpenAI warns that it may not include everything. ChatGPT also manages its saved memories on its own, keeping the most relevant ones at the front. Temporary chats don’t use memory and don’t create any.

Claude

According to Anthropic’s help center, memory from chat history is available to all users, from the free plan to Enterprise, on the web, Claude Desktop and the mobile apps. It is on by default on the Free, Pro and Max plans. On Team and Enterprise, organization owners decide whether it is available, and each member then has to turn it on.

Claude saves topics as the conversation goes, and you can tell it “remember this”: something you share today is known in the very next conversation. Each project has its own memory space, kept apart from the rest. Searching past conversations is limited to paid plans. You can see and edit what Claude remembers in Settings > Memory, and incognito chats stay out of it.

What these memories don’t do

These memories are handy so you don’t have to repeat your job, your style preferences or the names of your projects. They have three limits worth knowing:

  1. The assistant chooses what it keeps. A precise decision (“we’re keeping the short version of the tagline, approved by the client on the 12th”) may not be in there, or only in summarized form.
  2. They belong to each user. Memory is tied to your account. On Claude’s Team and Enterprise plans, owners cannot look at members’ memory. Your colleague gets nothing from what the assistant learned in your conversations.
  3. They don’t replace a reference document. You can’t read them through like a file, see who decided what, or when.

Custom instructions and Projects

Before memory, there were already two ways to give an assistant permanent context. They are still the most reliable, because you write the text yourself.

Custom instructions apply to all your conversations. In ChatGPT, they’re available on every plan, with a limit of 1,500 characters on Free and Go and 5,000 on the other plans. In Claude, you enter instructions in Settings and they apply to every conversation. Put in what doesn’t change: your role, how you work, the response format you expect.

Projects group conversations around one topic, with reference files and instructions specific to the project. They are the right tool for “one client” or “one campaign.”

Claude Projects ChatGPT Projects
Availability All plans, up to five projects on the free plan All plans
Contents Conversations, knowledge base (files), project instructions Conversations, files, project instructions
Memory A separate memory space for each project Option to limit memory to the project
Sharing With members of your organization, on Team and Enterprise Available on all plans according to OpenAI’s help center

A well-kept project solves most of the problem for one person: you put the brief, the brand guidelines and the decisions in the knowledge base, and every new conversation in the project starts with that context.

Five practical fixes

1. One context file per client or per job

Write a short document: who the client is, what they expect, the rules to follow, the decisions made (with their dates), the open questions. Add it to the project or paste it at the start of the conversation. People who use Claude Code do the same with a file read at the start of every session (CLAUDE.md). What matters is keeping it up to date: a wrong context file is worse than none.

2. Short, stable instructions

Keep general instructions for what applies everywhere (your job, your writing preferences), and put everything else in projects. Instructions that are too long or contradict each other get diluted.

3. One project per client

Many guides recommend it: one project per client or engagement, with its own files and instructions. One client’s conversations no longer get mixed up with another’s.

4. End each session with a recap

At the end of a session, ask the assistant to sum up what was decided and what’s left to do, then copy that summary into your context file. You stay in control of what gets remembered, instead of letting automatic memory choose.

5. An MCP memory server

The Model Context Protocol (MCP) lets you plug an assistant into external tools. Claude accepts custom connectors over remote MCP on all its plans (one custom connector on the free plan). An MCP memory server gives the assistant tools to write and read back information between conversations. The MCP project’s reference memory server, for example, stores a knowledge graph (entities, relations, observations) in a local file. It’s a good fit for one person on their own computer, less so for a team, since the file stays on a single machine.

With a team, the problem changes

On your own, you forget from one conversation to the next. As a team, everyone also needs to start from the same context. Take an agency where three people work for the same client: the project manager approved a change of tone on Monday, the copywriter opens Claude on Tuesday, the designer opens ChatGPT on Wednesday. None of their memories know about it, since each one is tied to its own account.

That raises several questions:

  • Where does the reference live? A shared project, a common document or a dedicated tool, but one place that everyone and every assistant reads.
  • Who can see what? A freelancer or a client should only see their own folder.
  • How do you track changes? Know who changed a rule, and be able to roll it back.
  • Where do you write down what’s left to do? Decisions and tasks go together: “we’re changing the tone” leads to “rework the three articles already written.”

Shared Projects answer some of these questions, within the limits of the plan and tool you pick. If your team uses several assistants, or you want to invite outside people into a single folder, a source of context that doesn’t depend on the assistant becomes useful.

Shared memory for a team: where Tavimi fits

Tavimi is the shared working memory for your team and your AI assistants, over MCP. It’s one option among others for teams, alongside shared Projects and context files. In practice, it’s an MCP server you connect Claude Code, Claude on claude.ai, Claude Desktop or any other MCP client to. Each client or job gets its own space, and an invited person only sees the space they’re invited to.

In each space, the team and its assistants write versioned notes (decisions, pitfalls, references), track tasks with an owner and a due date, and the assistant asks its questions on the task when it’s missing a decision, instead of guessing. Tavimi doesn’t replace your project management software or your assistant’s personal memory: it keeps what the whole team needs to know. To see how it’s organized, read how it works or the examples for marketing agencies.

FAQ

What people ask us

Is Claude's or ChatGPT's memory enough to keep track of a client?

It saves you from repeating your role or your preferences, but it keeps what the assistant judges useful, not necessarily the decision made last Tuesday. For a client, a written brief and a log of decisions, in a project or a document, are more reliable.

Do my colleagues benefit from what Claude remembered from my conversations?

No. Claude's memory, like ChatGPT's, is tied to your account. On Claude's Team and Enterprise plans, organization owners can turn the feature on or off, but they cannot see what each member's memory contains.

How do I start a conversation without the AI using what it remembers?

Claude offers incognito chats, which are not saved to your history and not used by memory. ChatGPT offers temporary chats, which don't use existing memories and don't create new ones.

Does one very long conversation solve the problem?

Not really. The context window has a size limit, and Anthropic's documentation notes that accuracy drops as the context grows. Short conversations work better, each one fed with written context that you keep up to date.

Give your team and your AI assistants one shared memory

Tavimi keeps the decisions, notes and tasks of every project, and Claude reads them back over MCP.

The Tavimi web app is in French for now; an English interface is on its way. Your assistants work in any language.