How do I know if my AI memory is actually broken?

Three quick tests. (1) 'What do you remember about me?' — should return a coherent short summary. (2) Reference something you know is in memory and see if the model uses it naturally. (3) Check the memory/settings panel and see if entries you expected are actually there. If any of the three fails, work through the diagnostic tree in this page.

What's the first thing to check when AI forgets you?

The Memory setting itself. Across every provider, the #1 cause of 'AI forgot me' is that Memory got toggled off — by a settings sync, an account switch, or a client update. Verify it's on before anything else.

How do I fix ChatGPT memory that stopped working?

Walk the 8-step playbook in our dedicated /chatgpt-memory-not-working-fix guide. The short version: check the setting, check for Temporary Chat, prune the memory cap, verify the account, refresh the client, rephrase to trigger retrieval, disable browser extensions, check for account-level pauses.

How do I fix Claude Projects context loss?

Verify the Project's system instructions are still populated, check that uploaded files are still attached, and confirm you're inside the Project (not in a standalone chat). If a specific Project stopped working after a model update, add a short voice-contract at the top of each new chat inside it.

How do I fix Gemini forgetting Workspace context?

Check Google Activity Controls — Personalized results must be on for Workspace data to be used. If it's on and Gemini still isn't pulling from Gmail/Drive, sign out of all Google accounts, sign back in with the one you actually want Gemini scoped to, then retry.

Is there one universal AI memory fix?

Yes: keep a canonical version of what the AI should know about you outside the AI, so any time provider memory breaks, you can paste your context back into any chat and keep going. Every fix on this page is patching a symptom; that's the durable prevention.

Encyclopedia · 25 min read

The AI memory troubleshooting encyclopedia

One reference for every way AI memory breaks, indexed by symptom and by vendor. Bookmark this. Send it to friends. It's the guide we wish existed the first time our own memory silently stopped working.

Tangled cables and gears resolving into orderly parallel diagnostic lines

Quick navigation

1. The universal diagnostic tree
2. Index by symptom
3. Index by vendor
4. Signals it's not fixable inside the provider
5. The structural prevention

1. The universal diagnostic tree

Regardless of the AI, ask these five questions in this order. Nine times out of ten you'll find the fix before you get to question 5.

Is memory turned on? Nearly every provider has a Memory setting that can be quietly toggled off by a settings sync, an account switch, or a client update.
Am I in an ephemeral mode? ChatGPT Temporary Chats, Claude "shared conversations," Gemini Incognito — all skip memory. Check the mode indicator before assuming memory is broken.
Is the memory store full or evicted? Every provider has a soft cap. Old memories quietly get overwritten by new ones with no notification. Open the memory panel and look.
Am I on the right account? Memory is per-account. Workspace SSO, secondary accounts, and Google-account switches can put you into a store with none of your memories.
Did the model just update? A model update changes retrieval behavior without touching stored memories. The panel looks identical, the model behaves differently.

If you get through all five without a fix, the failure is one of the harder categories: browser extension stripping headers, account flag pausing writes, or a genuine provider outage. See section 4.

2. Index by symptom

"It forgot everything about me"

Almost always memory setting off, wrong account, or Temporary/Incognito mode. Check in that order. If memories are actually gone from the panel, restore from export or move to a portable layer immediately.

"It remembers some things but not others"

Retrieval didn't fire for the missing bits. Rephrase your prompt with explicit reference ("based on what you know about me…"). If still missing after rephrasing, the entry has been evicted from the store. See memory limits for how eviction actually works.

"It's making things up about me"

Confabulation during retrieval. The model doesn't have what your prompt implies it should, so it fabricates a plausible answer. See how to fix wrong AI memory and when AI hallucinates about you.

"It feels different since the update"

Persona drift after a model release. Memories are intact, retrieval and generation have shifted. See AI personality changed after update.

"It forgot my character mid-roleplay"

Long chat pushed early context out of the window. Character.AI, Poe, and Replika are especially prone. Move the durable traits into the persona/definition field. See Character.AI memory guide.

"It contradicts itself between chats"

Different memories retrieved on different turns. See why AI memory contradicts itself.

"It remembers things I never told it"

Almost always Gemini or Copilot pulling from your Workspace/Microsoft-Graph data. Not memory in the saved-notes sense — live query at request time. If it feels invasive, adjust Personalized results in the account privacy settings.

3. Index by vendor

ChatGPT

Two independent switches (saved memory + chat history) — check both. Temporary Chat lightning-bolt is the sneakiest cause. Full 8-step playbook: /chatgpt-memory-not-working-fix.

Claude

No global memory layer — everything is scoped to a Project. If a Project "forgets," check the Project's system instructions and uploaded files. For cross-Project continuity, paste a shared context block at the top of each Project's instructions. Related: Claude Projects & model deprecation.

Gemini

Three layers to check: saved info panel, Personalized results toggle in Activity Controls, and the current chat's context. Full guide: Gemini memory guide.

Perplexity

Threads reset per session; Spaces persist context; Memory (Pro) crosses sessions. Most "memory broke" complaints on Perplexity are actually "I started a new Thread instead of continuing inside a Space." Guide: Perplexity memory guide.

Copilot

Split brain: consumer Copilot has its own memory; M365 Copilot draws from Microsoft Graph. Fixes for one don't apply to the other. Comparison: Copilot vs ChatGPT memory.

Grok

X integration means "remembers" is heavily influenced by your public profile. If Grok is bringing up things you posted years ago, that's the tie-in, not stored memories. Guide: Grok memory guide.

Character.AI

Three-layer memory (character definition + user persona + chat window). Almost every "the bot forgot me" complaint is fixable by promoting facts from the chat into the persona field. Guide: Character.AI memory guide.

4. Signals it's not fixable inside the provider

Some categories of failure aren't fixable by any settings change. Recognize them early so you don't spin.

The model was deprecated. Your saved memories may still be there, but the model that "knew" you the way you liked is gone. See the running deprecation list.
Account flag paused writes. New memories aren't being saved. Reads still work. Contact provider support.
Enterprise MDM policy strips headers. Memory works in personal browser, breaks on work laptop. IT decision, not yours.
Provider outage. Rare, and usually visible on status.provider.com. Wait it out.
Personality drift after model update. Not "broken" — different. See prevention below.

5. The structural prevention

Every fix on this page is patching a symptom of a system you don't own. The prevention that actually works is to keep the things that matter — your voice, your preferences, your projects, your key facts — in a place you control, and paste the relevant slice into any chat, on any provider, whenever it matters. That way "AI memory broke" becomes "AI memory doesn't apply anyway, I have my own layer."

That's the pattern Konshus is built around, and the deep version is in the complete guide to owning your AI consciousness. If you're here because something just broke, the fastest first step is to export what's still there while it's still there, and keep it somewhere the next model update can't touch.

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

Frequently Asked Questions

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus