As of May 2026. After 18 months of category-wide testing across every shipping AI memory product available to install, subscribe to, or evaluate: ChatGPT Memory, Claude's Memory tool, Gemini Past Chats, Mem, Reflect, Memorae, Saner, MyMind, Notion AI, Microsoft Recall on a Copilot+ laptop, Limitless's Pendant before Meta bought it, and Rewind on macOS before Meta sunset it. The hypothesis going in was simple: surely one of these already solves the problem of a phone that captures everything you actually need to remember, then answers questions about it in plain English. The honest answer at the end of 18 months is no, not one of them does the whole job, and that is why MemX exists. Here is the category-by-category review.
Quick takeaways: built-in chat memory (ChatGPT, Claude, Gemini) only remembers conversations you have inside that one chat box. Notes-vault apps (Mem, Reflect, Notion AI) only remember what you type. OS recall (Recall, Rewind) is desktop-locked and a privacy minefield. Wearables (Limitless, Pendant) are dead or dying. Nothing on the market does multi-modal phone capture (photos, PDFs, voice, chats) with on-device intelligence, per-user hardware encryption, and a single conversational query surface across models. That is the gap MemX is built for.
The five shapes AI memory comes in today
Before getting into the specific products, the category map matters. Every AI memory tool shipping right now falls into one of five shapes, and each shape is optimised for a different user. Knowing the shapes is the difference between picking a tool that fits your life and signing up for a $12 monthly subscription that ends up abandoned in week three.
- Model-built-in memory. Lives inside one chat product. ChatGPT, Claude, Gemini. Remembers what you typed in that product, nothing else.
- Notes vault with AI. You type or clip into a personal database; AI searches and synthesises. Mem, Reflect, Notion AI. Works if you live in the keyboard.
- WhatsApp or chat-native memory. You forward stuff to a bot; the bot stores and answers. Memorae, Saner.ai, MemX. Capture-first, low friction.
- OS or browser recall. Continuously screenshots your screen and indexes the text. Microsoft Recall, Screenpipe. Desktop only. Powerful, polarising on privacy.
- Wearables and ambient recorders. A device on your body captures audio or video. Limitless Pendant, Rewind, Friend. The most ambitious, the most fragile commercially.
Model-built-in memory: useful, narrow, model-locked
ChatGPT Memory
OpenAI launched memory as a limited test on February 13, 2024, then broadly to Free, Plus, Team, and Enterprise on September 5, 2024. The April 10, 2025 "reference chat history" update gave it the power to pull from any past chat, not just the saved-fact list. By 2026, memory is on by default for Free, Plus, Pro, and Team users; you can turn it off in settings or use Temporary Chat to bypass it for one session.
What it does well. Remembers your stated preferences (tone, formats you like, project context) so you stop re-explaining yourself at the top of every session. Pulls a relevant past conversation into a new one when the topic matches. Genuinely useful for the kind of person who works inside ChatGPT all day.
What it does not do. It cannot ingest a PDF you scanned three months ago on your phone. It cannot remember the prescription you photographed at the pharmacy. It cannot hold the voice note you recorded after a meeting, the screenshot of the rental agreement, or the WhatsApp message thread from your accountant. It is a chat memory, not a life memory. The saved-fact list is a soft word budget (community estimates put it at around 1,200 words on Free, 1,500-plus on Plus or Pro since the February 2025 +25% bump, which works out to roughly 100 to 200 short entries in practice). If you actually need to remember the world your phone has been recording, ChatGPT Memory is the wrong shape.
Claude Memory tool
Anthropic shipped the Memory tool in beta in August 2025 and brought it to free Claude users in March 2026. Memory is scoped per Project by default. Projects are intentionally isolated memory pools, which is a feature for privacy and a friction for portability. You cannot ask Claude in one Project what you saved into another, and you cannot ask Claude anything at all about a document that never entered Claude in the first place.
Gemini Past Chats and Personal Intelligence
Google rolled Past Chats to free Gemini users worldwide except Europe on February 26, 2026, then expanded Personal Intelligence (the broader Google-account-context layer) to free users in March 2026. Gemini's memory is account-bound, non-exportable, and lives inside the Google graph. If you also use Gmail and Calendar, the cross-product context is genuinely useful. If you want to take your memory with you to Claude or to your own AI agent, you cannot.
Common failure mode of all three. The memory is locked to the model. The day you want to switch from ChatGPT to Claude, or Claude to Gemini, none of it comes with you. Your life is captive in someone else's chat product.
Notes vault with AI: powerful for typists, hostile for everyone else
Mem.ai
Mem 2.0 shipped early 2026 and improved retrieval considerably. Pricing as of May 2026: free tier 25 new notes per month, Pro $12 per month. Works well if you already journal or take meeting notes by typing. The retrieval is fast and the AI summaries are good.
The catch is the input. Mem assumes you will type. It does not ingest photos of receipts, scan PDFs, transcribe voice notes that already live on your phone, or pull from WhatsApp threads. For someone whose life is more screenshots than journal entries, Mem becomes a vault that stays mostly empty.
Reflect
Reflect is a more polished notes vault, $10 per month with no free tier (14-day trial only). End-to-end encrypted notes, daily-note workflow, AI chat against your own writing. If you are already a daily-note person and willing to type, Reflect is one of the cleanest products in this space. Same fundamental limit as Mem: it is a vault for what you write, not for what your phone has captured.
Notion AI
Worth being precise here because the marketing is confusing. Plus at $10 per user per month annually includes basic AI writing only. Full AI Agents and Ask Notion (the actual cross-workspace memory layer most people picture) require Business at $20 per seat since the May 2025 restructure. So the real cost of "Notion as your second brain" is $20 per month per person, and it is designed for teams, not for individuals managing their own life.
Chat-native memory: closest in shape to MemX
Memorae
Memorae is the closest comp to MemX in capture pattern: WhatsApp-native, voice and text capture, semantic search. Pro tier $2.99 per month. If you are deciding between MemX and Memorae purely on "is there a WhatsApp memory bot," both qualify. The differences are in what is captured and how it is protected: MemX adds on-device photo classification (Google ML Kit, two-stage classifier so your selfies stay yours and only document-shaped photos are surfaced), Whisper transcription for voice up to a minute, scanned-PDF OCR with extracted entities (names, dates, IDs, amounts), and a Health Records feature that turns multi-year lab PDFs into trend queries ("what was my HbA1c over the last three years"). Memorae handles the chat surface well; MemX handles the chat surface and the rest of the phone.
Saner.ai
Saner.ai is for the Slack-and-Google-Drive professional. Starter $8 per month, Standard $16 per month. The integrations are genuinely useful for someone whose work lives across Slack channels and Drive folders. It is not built for a phone-first user capturing photos, prescriptions, and voice notes on the go.
OS recall: desktop-locked, privacy-loaded
Microsoft Recall hit general availability on Copilot+ PCs in April 2025. It screenshots your desktop continuously, runs OCR locally, indexes the result, and lets you ask natural-language questions about anything you saw. As a piece of engineering it is impressive. As a product it is locked to Windows on Copilot+ hardware, raised serious enough privacy concerns at launch that Microsoft delayed it and added opt-in plus per-app exclusions, and does not exist on the device most of your day actually happens on, which is the phone in your pocket.
Screenpipe is the open-source equivalent and avoids the Copilot+ lock-in. Same fundamental category: desktop-only, screenshot-driven, all-or-nothing capture. Genuinely useful for an engineer at a desk. Not the right shape for a parent at a paediatrician's office trying to remember the dosage from the last visit.
Wearables and ambient recorders: ambitious, fragile
Limitless (formerly Rewind) shipped the Pendant, a wearable that recorded ambient audio for ambient memory. Meta acquired the company on December 5, 2025, stopped Pendant sales, and sunset the Rewind desktop app on December 19, 2025. Friend, Bee, and a few other ambient-AI startups are still standing but the category has had a brutal 12 months. The lesson for buyers: do not hardware-lock your memory layer to a single startup. If the company gets acquired, you lose the device, the cloud, and access to whatever you already captured. A memory layer that is just an app talking to a backend you can swap is dramatically more resilient.
The gap none of them filled (and why MemX exists)
After 18 months of testing, the shape of the missing product became obvious. A real personal memory layer for an adult human in 2026 needs to do six things at once, and nothing on the market combines them.
| Capability | What every other tool does | What MemX does |
|---|---|---|
| Capture modes | Text only (chat or typed notes), or screen-only (Recall), or voice-only (wearables) | Photos, scanned PDFs, voice notes, chat messages, WhatsApp archives. Same app, same memory |
| Intelligence location | Mostly cloud; some local OCR on macOS Live Text or Recall | On-device first: Google ML Kit OCR, two-stage photo classifier, Silero VAD all run locally before any cloud call |
| Encryption model | Cloud-side at best; shared keys; vendor-readable in practice | SQLCipher AES-256 with a per-user key in Android Keystore / iOS Keychain; key never leaves the device; sign-out destroys it |
| Query surface | Locked to one model (ChatGPT, Claude, Gemini) or one vault | Ask MemX uses Vertex AI 768-dim embeddings + Firestore vector search, then synthesises with Claude or Gemini at your choice, with citations to the original source |
| Health records | Not a first-class feature in any general-purpose AI memory tool | Upload medical PDFs; multi-year trend queries ("HbA1c over three years") return a one-sentence answer with citations |
| Quick capture | Open app, navigate, paste, save | Persistent Android notification with theme-aware accent; one tap to record audio or save a document from anywhere on the phone |
The privacy story matters more than the feature list
Industry surveys put the share of consumers concerned about how companies use their data at around 80 percent, and three-quarters have stopped buying from vendors they do not trust. Health data, family photos, immigration documents, and financial records are not the categories where you accept a vague privacy policy and a cloud backup of unknown jurisdiction. The MemX design choice is to encrypt at rest on the device with a per-user key in the hardware-backed Keystore (Android) or Keychain (iOS). The key is generated on first sign-in, never leaves the device, and is atomically destroyed on sign-out. On a shared phone, User A's memories are cryptographically unreadable to User B even if they have the same app installed. That is not a marketing line; it is a SQLCipher AES-256 implementation detail.
The honest verdict by user type
- If you live in ChatGPT and only need it to remember your tone preferences, ChatGPT Memory is fine. Free, built-in, no extra app.
- If you live in Claude Projects and the per-project silos are a feature for you, Claude's Memory tool is the right pick.
- If you live in Google Workspace and want Gemini to know your Gmail and Calendar, Personal Intelligence is genuinely useful and now free.
- If you journal daily and want a beautifully designed vault, Reflect at $10/month is the cleanest. If you want a free starting point, Mem at 25 notes/month works.
- If your team is already on Notion, pay for Business ($20/seat) and use Notion AI. The Plus tier does not include the actual AI memory layer.
- If you want a WhatsApp memory bot for chats and short voice and that is the whole job, Memorae at $2.99/month is solid.
- If you want a phone-first memory layer that captures photos, scanned PDFs, voice notes, and chats; runs classification on-device; encrypts per-user in hardware; surfaces health records as a first-class feature; and lets you query through Claude, Gemini, or its own chat interchangeably, MemX is built for exactly that gap, and it is free to start.
What 18 months of building in this category surfaces
Three things stand out from the 18-month sweep. First, the chat-memory features built into ChatGPT, Claude, and Gemini are not in the same product category as the AI memory app market: they are features of one chat product, not a layer over a life. Treating them as competitors to a personal memory app is a category error. Second, the wearable bet was wrong for now: ambient hardware is a brittle business and Meta's acquisition of Limitless has effectively cleared the field of the most ambitious players. Third, the missing piece in 2026 is not better models. It is on-device intelligence that respects the phone (the device 80 percent of the world's adults use multiple hours a day), encrypts per user (because shared phones are still the global norm), and refuses the temptation to ship everything to the cloud just because cloud LLMs are cheap. The future of personal AI is not in your browser. It is in the device in your pocket, doing more work locally than the last generation of memory apps was willing to.
Try MemX free at memx.app. Scan five documents, record two voice notes, and ask MemX something about one of them a week later. That is the test that matters more than any feature comparison.
01Is MemX a notes app or an AI memory app?
An AI memory app. The difference matters. A notes app stores what you type. MemX stores what you live: photos of receipts, scanned PDFs, voice notes, chat messages, WhatsApp archives. Then it lets you query all of it in plain English, with citations to the original source. You do not have to type to remember.
02How is MemX different from ChatGPT Memory or Claude Memory?
Built-in chat memory only remembers what you said in that one chat product. It cannot ingest a PDF you scanned, a receipt you photographed, or a voice note you recorded. MemX captures all of that on the phone itself and makes it queryable from MemX or from Claude or from Gemini via MCP. Built-in memory is a feature of a chat app. MemX is a memory layer for your life.
03How is MemX different from Memorae or Saner.ai?
Memorae is the closest comp on the chat-native capture pattern. The differences are in capture breadth (MemX adds on-device photo classification, scanned-PDF OCR, voice transcription, and a Health Records mode) and in privacy posture (MemX encrypts at rest with a per-user key in hardware-backed Keystore or Keychain). Saner.ai is built around Slack and Google Drive integration, which is a different shape entirely.
04Is my data encrypted? Who can read it?
Every byte on the device is encrypted with SQLCipher AES-256 using a key generated per user on first sign-in and stored in the hardware-backed Android Keystore or iOS Keychain. The key never leaves the device. Signing out atomically destroys it. On a shared phone, User A's memories are cryptographically unreadable to User B even if they have the same app installed. We do not train models on your content. You can export and delete on demand.
05Does MemX work offline?
Capture, on-device OCR, photo classification, and voice activity detection all run locally on the phone (Google ML Kit, Silero VAD). You can capture documents and voice notes with no signal at all. Cloud-side synthesis (Ask MemX, RAG answers, long-form summaries) needs connectivity, but the capture path does not.
06Can I use MemX with multiple AI models?
Yes. MemX uses Google Vertex AI 768-dim embeddings stored in Firestore vector search, then synthesises with your choice of Claude or Gemini. Because the memory layer is model-agnostic, switching models does not lose your memory. That is the whole point of building a memory layer outside the model rather than inside any one chat product.
07Why is MemX phone-first instead of browser-first?
Because the phone is where the captures actually happen. The receipt you want to remember is a photo on your phone. The lab report is a PDF on your phone. The voice note from your contractor is on your phone. The WhatsApp thread from your accountant is on your phone. A browser-first memory tool optimises for a desk-bound knowledge worker. A phone-first memory tool optimises for the device a parent in Bengaluru, a digital nomad in Lisbon, and a sales rep in Chicago all hold for eight hours a day.
