DEV Community

How to Review, Edit and Delete What Your AI Coding Tools Remember About Your Project

Delete AI coding assistant memory is the act of permanently removing a stored decision, fix or preference from the memory your AI coding tools read, so that no future session is briefed with it. It is different from clearing a chat. Clearing a chat empties one conversation, while deleting a memory removes the record that would have been carried into the next one.

With Open Agentic Brain & Memory (OpenAB), the action itself is short. Open the local control panel on your machine, search for the memory, then edit or delete it. Because Claude Code, Cursor and Codex all read the same store, you do it once. The rest of this guide covers when to do it, how to find the right record, and how to check that the change took.

Why you sometimes need to delete AI coding assistant memory

Captured memory is written by a machine watching you work. Most of what it keeps is right. Some of it goes stale, and a small amount was wrong from the start.

Take a real pattern. You decide in Claude Code on Monday to drop Redis for SQLite. Three weeks later the queue outgrows SQLite and you bring Redis back. If the old decision is still sitting in memory, Cursor will confidently tell you on Tuesday that this project doesn't use Redis. The tool isn't broken. It is repeating what it was told.

There is also a cost argument. OpenAB injects a briefing of at most 2,000 tokens when a session starts, and that cap is the point. In our published benchmarks, loading context for a 10,000-memory store takes 1,325 tokens instead of the 1,131,367 it would take to load everything, and on LoCoMo (1,540 questions) OpenAB keeps 97% of the full-conversation score while reading 18% of the tokens. The benchmark numbers and how they were scored are on the site. A budget that small has no room for a stale decision, because every wrong line pushes out a right one.

OpenAB does some of this housekeeping without you. Memory is rewritten nightly and outcome scores let misleading memories sink. But a score can't know that you changed your mind in a meeting. At Open Agentic Brain & Memory (OpenAB), we think memory you can't read, edit and delete is memory you shouldn't trust, which is why every record is open to you in a local control panel.

A look once a month is plenty for most projects. Do it sooner after a big reversal, a migration, or any session where you pasted something you shouldn't have.

How to find the memory you want to delete: project, tool and date

Every memory carries three things that make it findable: the project it belongs to, the tool that recorded it, and when it was recorded.

Start with the project

OpenAB tags each memory by project, and it finds the project by walking up from your working folder to the nearest .git. So the first cut is nearly always the repo name. In the control panel, narrow to that project. From a terminal, the brain recall command does the same search over the same store, with a few words from the memory and a project name. Inside a session, a tool can run the brain_recall MCP tool with a query and a project. Recall over 10,000 memories takes 28 ms (p50) in our benchmarks, so there is no reason to guess when you can search.

Recall can search one project or all of them. Use the all-projects search when you suspect a preference leaked wider than it should, such as "always use pnpm" recorded in a repo that is still on npm.

Then use the tool and the date

Search uses both keywords and meaning, so "cache layer" will still surface the Redis decision even if those words never appeared in it. When you get several hits, the recording tool and the timestamp sort them out. If you remember making the call in Codex last Thursday afternoon, that is usually enough to pick the right record out of five similar ones.

Read the whole memory before you touch it. A distilled decision often carries the reason alongside it, and the reason may still be true even when the decision isn't.

Edit a wrong decision or delete it outright?

Our view is that you should edit more often than you delete.

Edit when the decision changed. "Dropped Redis for SQLite" becomes "Went back to Redis for the job queue after SQLite hit write contention; SQLite stays for local dev." That sentence is worth more than an empty slot. It stops the next tool from proposing the SQLite move again, which is exactly what a helpful assistant with no history will do.

Delete when the memory was never true, when it belongs to a project that is gone, or when it contains something that should not be stored at all. A misheard preference, a debugging dead end recorded as a fix, a pasted credential. None of these improve with rewording.

Pin is the third option, and it works the opposite way. If a memory is correct and you want it to stay in front of every tool, pin it instead of hoping the ranking keeps it there.

How to delete a memory, step by step

  1. Open the OpenAB control panel on the machine where you work. It runs locally, and there is no account to sign in to.

  2. Narrow to the project, then search with a few words from the memory.

  3. Open the record and read it in full. Check the tool and date to confirm it is the one you mean.

  4. Delete it, or edit it if the decision changed and the history is useful.

  5. Run brain sync so the encrypted copy in your private GitHub repo is replaced with the current state.

  6. Start a new session in each tool you use and confirm the memory is gone, as described below.

Checking that the deleted memory is gone from Claude Code, Cursor and Codex

There is one store, a SQLite database on your machine, and every connected tool reads from it. You are not chasing three copies. The list of tools OpenAB connects to also covers Antigravity and Kilo Code (beta), and the same check applies to those.

The one thing that catches people out is open sessions. The briefing is injected by a start-up hook when a session begins (84 ms at p50 in our benchmarks), so a chat that was already open still has the old briefing in its context. Close it and start a new one.

Then check two ways:

  • Ask the plain question. "What do we use for the job queue in this project?" The answer should reflect your edit, or admit it doesn't know if you deleted.

  • Ask the tool to search memory for the old wording. An empty result from brain_recall is the confirmation you want.

Do this in each tool once. It takes a couple of minutes, and it is the only way to know instead of assume.

FAQ

Does clearing a chat delete AI coding assistant memory?

No. Clearing a chat removes that conversation's context and nothing else. Anything already captured as a memory stays in the store and will be briefed into the next session until you edit or delete it in the control panel.

How do I delete a memory in Claude Code, Cursor or Codex?

With OpenAB you don't delete it inside each tool. You delete it once in the local control panel, and since all three tools read the same store, the memory is gone for all of them. Start a new session in each tool so it picks up a fresh briefing.

If I delete a memory on one laptop, is it removed on my other machine too?

Not at that instant. The delete happens in the local database on the laptop you are using, and sync is daily and sends one whole encrypted file. Run brain sync on the machine where you made the change, then open the control panel on the other machine after its next start-up and confirm the memory is gone. If it is still there, delete it there too.

What should I do if a secret or API key ends up in my coding agent's memory?

Rotate the key first. The secret has already passed through a chat session, so deleting the memory alone does not make it safe again. Then delete the memory in the control panel, run brain sync to replace the synced file, and check your other machines. The synced copy is age-encrypted with a key kept only in your OS Keychain, so a leaked repo or token exposes nothing readable, but that is a reason to worry less, not a reason to skip rotation.

Can a deleted memory be recovered?

Treat a delete as final. A copy may still exist on another machine or in the last synced file until the next sync replaces it, but don't plan around that. If you are unsure, edit the memory instead of deleting it.

Where are the memories stored?

In a SQLite database on your own machine. For sync, the whole store is age-encrypted and saved as a Release asset on a private GitHub repo that you own. There is no account and no server of ours, and only you hold the key. If you lose it, nobody can recover your memory.

OpenAB is free and open source under Apache-2.0, built by Rohith, Chandan and Sushmita. If you want one memory across your tools that you can read and correct yourself, Install Open Agentic Brain & Memory (OpenAB).

Top comments (0)