Common problems (FAQ)

When a memory you expected isn't there, when saving doesn't happen, when search comes back wrong. What to check, in order.

I saved something but the memory isn't there

Before assuming data loss, check in order.

Is the save target right? Use the check-target tool to see the current account, team, and project. You may be on a different account in another browser session, or switched to a different project. The target has to match between when you saved and when you search.
Is it inside the search scope? The server sets the scope ("what I can read within my team" is the ceiling). Look at searched_scope in the response → Searching.
Did it actually save? If you spoke without the ak command, it did not save → Saving.

Most of the time it isn't loss, it's a target mismatch → Teams and projects.

I get a "temporarily unavailable / connection reset" error

This can be a cold start on the serverless database (the first attach after it's been idle). Call again a few seconds later and it reconnects automatically. If it repeats, retry after a while.

I removed the plugin in ChatGPT, installed it again, and the tools are gone

The old connection state can be cached, so the newly installed plugin doesn't attach. Check in order.

Fully restart ChatGPT. Don't just close the window. Quit the app and reopen it. There have been real cases of a failed-authentication state being cached so new values are never retried; a restart, or removing the plugin and installing it again, clears it.
Check that you turned the plugin on in the new conversation. ChatGPT enables plugins per conversation. Under Plugins below a new chat's composer, turn AiAkiv Memory on, or start from the Try now button shown right after installing. Even when it's saved in settings, a new conversation starts with it off (step 4 of §1 in Connecting ChatGPT).
If that doesn't work, check .codex/config.toml. The ChatGPT app inherits the plugins installed on your account as they are (codex_apps), and removing and installing again can leave a disabled state in the local config.
- First ask ChatGPT for the app id: "tell me this app's app id".
- Open .codex/config.toml in your user folder (e.g. ~/.codex/config.toml, or C:\Users\<name>\.codex\config.toml on Windows).
- Check whether the entry for that app id is turned off with enabled = false. If it is, change it to true or delete the line and restart.

A related caution. Keeping a local MCP config ([mcp_servers]) in a folder alongside it is fine in itself. But its name must differ from the account-wide connection's (the plugin install): global is AiAkiv, folder is AiAkiv-<project name> (e.g. AiAkiv-k2g). Copy it from the console or download the folder ZIP and it comes with that name. With identical names there's no way to tell which one was used → Connecting ChatGPT. Claude is different. Even with different names the AI can use the account connector. Still, leave the account connector on (turning it off removes it from the web and the mobile app too) and check where each save went → When the account connector and a folder connection are both attached.

It won't save

AiAkiv saves only on an explicit ak command. Just saying "save this" or "remember this" does not save (that's the guard against accidental saves). Try again with "ak save this: ..." → Saving.

The AI refuses, saying "Main is turned off"

Main is turned off in the console for this account. Nothing is broken. While it's off, connections that follow Main (the account connector in claude.ai web chat, the mobile app and Desktop, and API keys not pinned to a project) refuse saving, searching and listing alike.

If this conversation also has a folder connection (Claude Code, Desktop's </> Code), tell the AI to use that one. Folder connections and project-pinned API keys keep working while Main is off.
If you meant to use it on your phone or the web, press "Turn Main on" in the console's Projects → Project management. The refusal message the AI passes on includes a link to that screen. Turning it on brings back the Main you had just before, unchanged.
Asking the AI that refused to turn it on doesn't work. It is on a connection that follows Main, so every request from it is blocked. If "Turn Main back on" is on in AI permissions, you can ask an AI connected through a folder or a project-pinned API key to turn it on. On your phone, use the console button.

→ When the account connector and a folder connection are both attached.

I want to change the save target

A person changes it, in the console. The save, search and check-target tools cannot change the target (Main). That's the anti-hijacking guard. Click the switch_url given by the check-target tool or the list-projects tool. If you turned on the save target switch permission in the console's AI permissions, you can also ask the AI to switch it (Asking your AI to do console work). A connection made with an API key also needs that key to carry the manage scope (Authentication). Asking the AI works only on a connection that has app features; on one that does not, you switch it in the console yourself (The app action tool). → Teams and projects.

Search results look wrong or out of date

Don't take the summary at face value. It's LLM-generated and occasionally inaccurate. Confirm exact content with fetch full text.
Follow the connections. Reading a hit's reason and the link map (hint) can turn up a renamed synonym or a continuing fragment → Searching.
Judge conflicting information by recency.

The same thing comes back as two separate entries

The entity name most likely wobbled ("save format" vs "savefile format"). Save following the one-name-per-entity rule and they weave into one → Saving.

I want to fix something I saved

Don't re-save: you'll end up with two copies of the same content. There are two handles.

The entities, tags, or chaining are wrong → "ak fix the tags on what I just saved". The original text can't be changed: if you need to change that, retire it and save again.
You'd rather that memory stopped coming up at all → "ak retire what I just saved".

Details in Correcting a save afterwards in Saving.

I saved personal data or a secret by mistake

Not putting it in is the best answer (the do-not-store list in What to store).

If it's already saved, retire it to take it out of recall ("ak retire what I just saved"). But there's something you need to know exactly.

Retiring is not deleting. It drops out of search and recall, but the data remains. And there is no way for a user to physically delete a single memory , the only things that remove the original are deleting the team and closing your account (console → Danger zone).

So if the value is one you're worried about leaking, don't stop at retiring. invalidate the value itself. Re-issue an API key or token at its source; change a password. A secret being hidden from memory and that secret still being valid are two different problems.

Learn more

Try it → tour
Per-client connection recipes → playbook
When a term is unclear → console Reference → Help. Type a word and it filters immediately (Console)