The best AI note-taking apps, judged on whether they can find what you wrote

Prices and features checked: 2026-11-09. Prices in the currency the vendor shows, checked on the date shown; we do not convert.

What AI adds to notes, and what it breaks

Every note app now has an AI button. Most of them do the same four things: summarise, rewrite, transcribe and answer questions over what you have saved. Only the last one changes what a note system is for. Writing a note is cognitive offloading, a bet that you will be able to find the idea later without holding it in your head; a 2016 review concludes the bet usually pays, at the price of weaker unaided memory for whatever you filed.^[1] A Retrieval is where that bet is settled, and before AI it depended on you remembering a title, a tag or a folder. Chat over your own notes removes that requirement. That is the one genuine advance in this category.

It also breaks something. The learning research is consistent that practice testing and spaced retrieval are high-utility and rereading and highlighting are low-utility.^[2] A A model that summarises a note before you reread it saves time and removes the retrieval attempt that would have strengthened the memory. Summaries are a gain when the note is reference material and a loss when the note is something you are trying to learn. Keep that distinction in mind through every pitch below.

The productivity evidence is strong on bounded tasks and silent on note-taking. Professionals using ChatGPT finished writing tasks about 40 percent faster at higher rated quality, weaker writers gaining most.^[3] A An AI assistant raised customer-support output by about 14 percent, concentrated among novices.^[4] A Yet across 106 experiments, human-AI combinations on average did worse than the better of human or AI alone, except on content creation.^[5] A Automation bias, accepting the machine's confident answer unchecked, is well documented and is reduced by training and accountability, not by better prompts.^[6] A Reports that AI erodes critical thinking are cross-sectional surveys and small preprints and cannot show causation.^[7] C Expect time savings on summarising; expect no evidence yet that AI notes make you understand more; check the citation before trusting the answer.

Criteria

We judged all eight on the same five things, in this order. Retrieval design: does the tool answer from your notes with citations you can click, and what does it index? Meeting capture: can it turn a call into searchable notes, and at what cost? Privacy and training terms: what the vendor's own policy says, quoted and dated. Ownership: export formats, offline behaviour and whether anyone but you can read the notes. Price as shown on the vendor's page or in our catalogue check on 2026-11-09. Affiliate status did not affect the order: Notion's programme is closed to new applicants, Mem, Reflect, Google and Microsoft pay nothing, and Evernote, Capacities, Fathom and Readwise run programmes.

Demand is steady but not growing: US searches for "ai note-taking app" and its spelling variants run at about 3,600 a month each with a keyword difficulty of 17 to 23, down 32 to 45 percent year on year, while "note taking app free" is up 241 percent.^[8] People are asking less often whether notes should have AI and more often what it costs.

The retrieval benchmark: 50 questions over the same notes

Nobody in the current search results tests retrieval; they list features. We have not published results yet either, and we will not invent them. What we can give you is the protocol, so you can run it on your own notes in an afternoon and so our scores, when we publish them, are reproducible.

1. Corpus. The same 200 notes imported into every app: 100 personal notes of mixed length, 50 meeting notes, 30 web clips, 20 PDFs. Identical files, identical titles.
2. Questions. Fifty, written before testing, in five types of ten: exact fact ("what date did we agree with the printer"), paraphrase ("the argument about pricing tiers"), cross-note synthesis ("what did three meetings say about hiring"), negative ("did we ever discuss Lisbon", answer: no), and stale ("the latest version of the budget", where two notes conflict).
3. Scoring. Three points per question: correct answer, correct source note cited, no fabricated detail. Maximum 150. Negative questions score only if the tool says it cannot find the topic.
4. Controls. Plain keyword search in the same app on the same questions, so we measure what AI adds over search rather than how good the app's search is.
5. Report. Score by question type, time per answer, and the number of fabricated details, which matters more than the total.

What the designs predict, before any test: source-bound tools that cite (NotebookLM, Copilot Notebooks) should score highest on fact and synthesis questions and worst on stale ones, because they do not know which note is newer. Workspace-wide assistants (Notion, Evernote, Mem, Capacities) should do better on stale questions and worse on negatives, because a large index tempts a model to answer. Reflect cannot be benchmarked this way: its vault is encrypted and the AI sees only text you select.

Meeting capture

If most of your notes arrive through calls, the recorder is the note app. Fathom records and transcribes without limit on its free plan and charges from USD 16 a month billed annually for AI summaries and features; Team seats are USD 15 annual with a two-user minimum.^[9] Notion includes AI Meeting Notes on Business, which keeps transcripts next to the project they concern.^[10] Evernote offers AI Transcribe for recordings you make yourself.^[11] Capacities and Mem capture voice notes rather than meetings. NotebookLM and Copilot Notebooks summarise recordings you upload or that Teams produces, respectively.

The decisive difference is the training clause below, because meeting audio contains other people's words.

Summaries: how to check accuracy yourself

We will not quote an accuracy percentage we have not measured. Two five-minute checks catch most failures. First, ask for the summary, then ask the same tool for the three sentences in the source that support each claim; a summary that cannot be traced is a paraphrase of the model's expectations, not of your note. Second, plant one false but plausible detail in a copy of a note and see whether the summary repeats it, omits it or flags the conflict with the original. The jagged-frontier result from a field experiment with 758 consultants is the pattern to expect: strong inside the model's competence, confidently wrong just outside it.^[12] B

Privacy and training terms, quoted

Each statement is the vendor's own text, read on 2026-11-09, with the policy's date where shown. Vendors change these pages; re-read before uploading anything you could not bear to have trained on.

Tool	Training on your content	Who processes it	Retention with providers
Notion AI	"By default, your customer data is not used to train models"; subprocessors contractually barred	Mix of LLM providers listed on a subprocessor page	30 days or fewer on Free, Plus, Business; zero retention on most Enterprise models [10]
Evernote	"We do not use your Input or Output to train our artificial intelligence models or tools, or those of third-party AI model providers" (policy of 1 October 2026)	Unnamed third-party AI providers	Not stated [11]
Mem	Not used "to develop, improve, or train any generalized AI and/or ML models" (policy of 13 August 2026)	Subprocessors at trust.mem.ai	Not stated [13]
Reflect	"OpenAI and Anthropic may use prompts we send them as training data" (policy of 20 March 2025); only highlighted text is sent, unlinked to your account	OpenAI, Anthropic, Google	Audio path: provider deletes logs in about 30 days [14]
NotebookLM	"Your data is protected, and is not used to train Gemini Notebook unless you provide feedback"	Google	Not stated [15]
Copilot Notebooks	"Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs" (updated 4 November 2026)	Microsoft, with OpenAI and Anthropic models	Not stated [16]
Capacities	"No training on your data – your content is not used to train models"; DPAs prohibit it	Anthropic (text), Google Gemini (media); picker includes OpenAI, xAI, Mistral, DeepSeek, Moonshot	Up to 30 days for abuse monitoring [17]
Fathom	Fathom "may use and create de-identified data generated from Meeting Content Information" to train its in-house models; opt-out in settings; third parties may not (policy of 16 August 2026)	Fathom; unnamed providers	Recordings deleted within 30 days of account deletion [9]

Two of the eight stand out. Reflect is the only one honest enough to say its providers may train on what you send, and it limits what you send to text you choose. Fathom is the only one that reserves the right to train its own models on your meetings, de-identified, unless you opt out; given that meetings contain other people's words, check that box before the first call.

Prices, USD and EUR

Prices as shown on 2026-11-09; EUR only where the vendor served it. "See pricing" marks pages we could not read.

Tool	Free tier	Entry paid plan	AI tier or add-on
Notion AI	Free with limited AI trial	Plus EUR 9.50 per seat as served (USD 10 annual, 12 monthly)	Business EUR 19.50 (USD 20 annual) includes Agent, Meeting Notes, Enterprise Search; Custom Agents USD 10 per 1,000 credits [10]
Evernote	50 notes, 1 notebook, 1 device	Starter USD 8.25 annual, 14.99 monthly	Advanced USD 20.83 annual, 24.99 monthly; AI features by plan not clearly shown [11]
Mem	25 messages, 25 notes a month, 3 hours voice	Plus USD 9 monthly	Pro USD 29; Pro 2x to 20x USD 49 to 199 monthly [13]
Reflect	Not shown	See pricing (page returned 404)	Included [14]
NotebookLM	Standard: 50 sources per notebook, 100 notebooks	Via Google AI Plus, Pro, Ultra; amounts not shown on the plans page, see pricing	Pro 300 sources per notebook, 500 notebooks [15]
Copilot Notebooks	Limited in Copilot Chat	Copilot Business add-on USD 21 list, 18 promotional, per user per month yearly; 25.20 monthly	Microsoft 365 Copilot add-on USD 30 per user per month paid yearly, plus a qualifying Microsoft 365 licence [16]
Capacities	Basic, unlimited objects	Pro USD 9.99 a month on the page (structured data: 11.99 monthly, 119.88 yearly)	AI Power-Up USD 8.33 a month billed annually [17]
Fathom	Unlimited recordings and transcripts	Premium USD 16 annual, 20 monthly	Team USD 15 annual, 19 monthly, 2-user minimum; Business USD 25 annual, 34 monthly [9]

Two tools sit outside the table. Heptabase meters AI at USD 8.99 a month for 100 credits and USD 17.99 for 1,800 and states "We do not use Your Content to train AI models".^[18] Tana, now a meeting-notes product at USD 15 per member per month billed yearly, says in its FAQ that data is never used to train models while its privacy policy of 18 May 2026 promises only "advance notice and an opportunity to opt out"; we quote both because they do not match.^[18] Our PKM guide covers both. Obsidian has no AI of its own, but the Smart Connections plugin builds embeddings locally with no API key, the only route on this page with no vendor policy to read; Obsidian vs Notion sets out that trade.

Verdict

• Best overall for an existing workspace: Notion AI on Business, because retrieval runs over the notes, docs and meetings you already keep there, and the training terms are clear. The cost is the plan jump.
• Best for research from documents: NotebookLM. Grounded, cited answers over a fixed set of sources, free at 50 sources per notebook. It is not a notebook for your own thinking.
• Best privacy: Reflect, with its eyes open: the vault is end-to-end encrypted, the AI sees only what you select, and the policy tells you the providers may train on that selection.
• Best transparency on providers: Capacities, which names the model behind each feature and the retention period.
• Best for meetings: Fathom on price and limits, with the opt-out ticked.
• Best for a large legacy archive: Evernote, whose semantic search gives an old archive a second life and whose policy rules out training.
• For Microsoft shops: Copilot Notebooks, if the organisation already pays for Copilot; it is not a consumer purchase.

Who should not buy

• Do not buy an AI tier for notes you are trying to learn; the summary removes the retrieval attempt that makes the memory stick. Use AI for reference material and meetings.
• Do not upload client or patient notes to any tool whose policy you have not read this month; the table above is a snapshot.
• Do not buy Notion Plus for the AI; the useful features start at Business.
• Do not pick Mem or Reflect if you need a team workspace; both are personal tools.
• Do not pick NotebookLM as your note app; it answers over sources, it does not hold a daily note.
• Do not record meetings in Fathom without checking the training opt-out and telling participants.

Decision table by user type

You are	Pick	Why	Monthly cost, annual billing
Team already in Notion	Notion Business	Retrieval over everything you keep there	EUR 19.50 per seat as served
Student or researcher with a reading list	NotebookLM Standard	Cited answers over 50 sources, free	USD 0
Privacy-first individual	Reflect, or Obsidian with local embeddings	E2E encryption; AI only on selected text	See pricing / USD 0
Someone who wants to chat with a messy pile of notes	Mem Plus	Chat-first design, Markdown export	USD 9 monthly
Person with ten years in Evernote	Evernote Starter or Advanced	Semantic search over the archive	USD 8.25 or 20.83
Knowledge that arrives through calls	Fathom Free or Premium	Unlimited transcripts; opt out of training	USD 0 or 16
Structure without databases, clear AI provenance	Capacities Pro plus AI Power-Up	Named providers, 30-day retention	USD 9.99 plus 8.33
Microsoft 365 organisation	Copilot Notebooks	Grounded in your tenant, no training on Graph data	USD 30 per user with Copilot

What to do next

Run the ten-question version of the benchmark above on your own notes in two candidates before you pay for either: five fact questions, three synthesis, two negatives. Count fabricated details. Then read the training clause on the vendor's page that day, not ours. If you want the shortlist built from your devices, budget and privacy needs, the Tool Finder does that against these dated prices, and what a second brain is explains where AI retrieval fits in a complete note system. Fathom records for free without a card, and Evernote offers seven-day trials of its paid plans.

Frequently asked questions

What is the best AI note-taking app?

For a team already in Notion, Notion AI on the Business plan (EUR 19.50 per seat per month as served on 2026-11-09), because it retrieves over the notes and meetings you already keep there and the training terms are clear. For research from documents, Google's NotebookLM is free at 50 sources per notebook. For privacy, Reflect encrypts the vault end to end and sends only text you select to the model.

Do AI note apps train their models on my notes?

Most say no in writing: Notion ('By default, your customer data is not used to train models'), Evernote, Mem, Capacities, NotebookLM and Microsoft Copilot all exclude training on customer content. Two differ. Reflect states its providers, OpenAI and Anthropic, may use prompts as training data, though only highlighted text is sent. Fathom may train its own models on de-identified meeting content unless you opt out in settings. All read on 2026-11-09.

Is there a free AI note-taking app?

Yes. NotebookLM's Standard tier is free with 50 sources per notebook and 100 notebooks. Fathom records and transcribes unlimited meetings free. Mem's free tier allows 25 messages and 25 new notes a month. Notion's Free plan includes a limited AI trial. Obsidian with the Smart Connections plugin gives local AI search free with no account (checked 2026-11-09).

Does AI help you remember your notes?

It helps you find them, which is different. Offloading to notes improves task performance but lowers unaided recall, and the learning research rates retrieval practice as high-utility and rereading summaries as low-utility. A Use AI retrieval for reference material and meetings; for things you need to know without looking them up, keep testing yourself.

Notion AI or NotebookLM?

Different jobs. Notion AI answers over your whole workspace and writes meeting notes, from the Business plan. NotebookLM answers over a fixed set of documents you upload, with citations, and is free for 50 sources per notebook; it does not hold your daily notes. Researchers often use NotebookLM to read and Notion or Obsidian to write.

Sources and evidence

1. Cognitive Offloading — Risko EF, Gilbert SJ (2016), review A doi.org/10.1016/j.tics.2016.07.002 (retrieved 2026-11-09)
2. Improving Students' Learning With Effective Learning Techniques — Dunlosky J, Rawson KA, Marsh EJ, Nathan MJ, Willingham DT (2013), systematic review A doi.org/10.1177/1529100612453266 (retrieved 2026-11-09)
3. Experimental evidence on the productivity effects of generative artificial intelligence — Noy S, Zhang W (2023), RCT A doi.org/10.1126/science.adh2586 (retrieved 2026-11-09)
4. Generative AI at Work — Brynjolfsson E, Li D, Raymond L (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
5. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro M, Almaatouq A, Malone T (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
6. Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard K, Roudsari A, Wyatt JC (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
7. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking — Gerlich M (2025), observational C doi.org/10.3390/soc15010006 (retrieved 2026-11-09)
8. Productivity Signal keyword database, US/en rows 'ai note-taking app', 'ai note taking app', 'ai-powered note-taking app', 'note taking app free' (DataForSEO Labs keyword_suggestions, location 2840) — Productivity Signal research, via DataForSEO (2026), dataset A dataforseo.com/ (retrieved 2026-11-09)
9. Fathom pricing and privacy policy (last updated 16 August 2026) — Fathom Video Inc. (2026), vendor page A fathom.ai/pricing (retrieved 2026-11-09)
10. Notion pricing (served in EUR), Notion AI product page and Notion AI security and privacy practices — Notion Labs (2026), vendor page A www.notion.com/help/notion-ai-security-practices (retrieved 2026-11-09)
11. Evernote compare plans and privacy policy (last updated 1 October 2026); prices from Productivity Signal catalogue check 2026-11-09 — Evernote (Bending Spoons) (2026), vendor page A evernote.com/compare-plans (retrieved 2026-11-09)
12. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua F, McFowland E, Mollick ER, et al. (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
13. Mem pricing and privacy policy (revised 13 August 2026) — Mem Labs (2026), vendor page A get.mem.ai/pricing (retrieved 2026-11-09)
14. Reflect privacy policy (last updated 20 March 2025) — Reflect App LLC (2025), vendor page A reflect.app/privacy (retrieved 2026-11-09)
15. Gemini Notebook (NotebookLM) help: plans and limits — Google (2026), vendor page A support.google.com/gemininotebook/answer/16213268 (retrieved 2026-11-09)
16. Microsoft 365 Copilot pricing (enterprise and business) and Microsoft Learn: Data, privacy, and security for Microsoft 365 Copilot (updated 4 November 2026) — Microsoft (2026), vendor page A learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot- (retrieved 2026-11-09)
17. Capacities pricing, AI assistant reference and AI privacy documentation — Capacities (2026), vendor page A capacities.io/pricing (retrieved 2026-11-09)
18. Heptabase pricing and privacy policy (updated 28 October 2026); Tana pricing and privacy policy (updated 18 May 2026) — Heptabase; Tana Inc. (2026), vendor page A heptabase.com/pricing (retrieved 2026-11-09)