MemoryBrain
How do I stop AI from using an old decision after I change it?
Date each decision, give it a scope and say what it replaces. Keep one current-state document and separate the history. Correct or delete outdated memories, then check in a new chat which decision it uses and its source. This is a method, not a guarantee.
Make the current version easy to check
Write the new decision with its date and scope
Say what changed, when it takes effect and which project it applies to. Fictional example: “November 8, 2026. Website project. Launch moves to November 20. Replaces the November 2 decision to launch on the 12th.”
Mark the superseded decision
Add “Historical. Superseded by the November 8 decision” to the old entry. Keep its reason and source if you need to explain the change. A more recent import date does not prove a decision is newer.
Update one state document
Put the current decision, next step and source there. Move previous changes to a separate history. Avoid several files called “current state” with different versions.
Review memories that contradict the state
If ChatGPT retains the old decision, inspect its memory controls and correct or delete the entry. In Claude, review and edit its memory. Also review documents you still use as context.
Ask from a new chat without supplying the answer
Provide the current document and ask it to identify the applicable decision, date and superseded decision. Ask about history separately. If it mixes versions, correct the sources before continuing work.
“What launch date is current? Cite the document and decision date. If you find another version, explain which one replaces which. Do not choose solely by file upload date.”
What the reported cases show
Thread 1405494 describes information apparently reverting to an older state and mixing with recent context. The author cannot determine the internal cause. In thread 1396965, another user requests access to history and later decisions that superseded earlier ones.
These are reports and a feature request. They do not establish a general outage, a single cause or that this procedure eliminates every error. The proposed method makes the applicable version explicit.
A template for state and another for history
- Current state: project, applicable decision, effective date, next step and source link.
- History: date of each change, old decision, new decision, reason and what it replaces.
- Awaiting confirmation: alternatives and unapproved proposals. Keep these separate from agreed decisions.
If a change only affects part of the work, write that exception. “The mobile version keeps the previous plan” is more precise than “forget everything before this.”
Correcting memory and preserving history are separate tasks
To stop treating a decision as current, mark it as historical and update the state. To remove a fact completely, also review its memories and sources through the service’s controls. Adding an instruction to “ignore the old version” does not prove the stored information has been corrected.
See how to correct a memory without adding duplicates and how to continue the project in another chat.
Where MemoryBrain fits
In MemoryBrain projects, the state replaces the previous one. You can save the next step and tasks, and each progress update or decision goes into a dated history. This separates where you are from how you got there.
When a decision changes, update the state and record what it replaces and why in the history. Then check the answer and source from ChatGPT or Claude. A dated history alone does not guarantee every answer selects the correct version.
Frequently asked questions
Is adding a date to the decision enough?
No. Add the project or scope and which decision it replaces. Keep one current-state version and check the answer against its source.
Must I delete every old decision?
Not to preserve project history. You can mark old decisions as historical and separate them from current state. To remove a fact completely, also review its memories and sources.
Does MemoryBrain guarantee AI will never use the old decision?
No. Project state replaces the previous state and decisions go into a dated history. You should still check which version the assistant retrieves and which source it cites.
Try it
See how a memory that keeps the source of everything works, no account needed.
Sources
- OpenAI forum: User Knowledge Memories reverting to older state (1405494). October 12, 2026. Accessed November 8, 2026.
- OpenAI forum: User-Initiated Conversation History Retrieval (1396965). September 12, 2026. Accessed November 8, 2026.
- OpenAI: Memory in ChatGPT. no absolute publication date stated. Accessed November 8, 2026.
- Anthropic: Claude’s chat search and memory. no absolute publication date stated. Accessed November 8, 2026.