ChatGPT Memory and Long-Term Projects: Make It Know You Better Over Time

Separate work with projects, set rules with custom instructions, replace endless chats with a progress file, and prune memory regularly.

After a while, ChatGPT tends to remember the wrong things and forget the right ones. The fix is one idea: keep long-term information somewhere you can see and edit, instead of relying on the model to remember it.

Three layers, three jobs

Custom instructions: things that never change, such as who you are, answer style and hard rules. Keep them short.

Memory: personal preferences across chats. It fills up on its own, mistakes included.

Projects: files, rules and chats for one specific job. Anything you work on for weeks deserves its own project.

Setting up a long-term project

One job per project, for example "store marketing", "exam prep" or "weekly report". Do not mix them.

Project instructions cover three things only: goal, constraints, output format.

Upload key material as project files. It is more reliable than pasting it into chats again and again.

Turn on project-only memory when you do not want other chats to leak into the project.

Use a progress file, not an endless chat

The longer a chat gets, the more its early parts get ignored. At the end of each stage, ask for a summary in a fixed format and save it to the project files. Start new chats from that file instead of continuing one huge thread.

Try: "Summarize the decisions made in this chat, the open questions and the next steps in under 200 words."

Prune memory regularly

Review the memory list once a month and delete outdated, wrong or one-off entries.

If it keeps getting a fact wrong, check whether memory stored it wrong.

Use a temporary chat for questions you do not want remembered.

Never let it remember passwords, ID numbers or card details.

Common mistakes

Instructions written as an essay: keep the 5 rules that matter most.

Everything in one chat: the context gets messy and answers drift.

Memory without files: nothing comes with you when you switch accounts or tools. See keeping memory across AI tools.