Everything you need to get the most out of MemoRoot.
Two things are required before Capture will work: an API key from a supported LLM provider and a connected vault (GitHub or Google Drive). If either is missing, Capture shows a message prompting you to connect it - nothing is saved until both are in place.
MemoRoot uses your own API key for note organization and answers, so you pay only for what you use and your key stays under your control. (Two AI features run on MemoRoot's own keys instead: search embeddings via Voyage AI, and the product Help widget.) The steps below use Anthropic, the recommended provider; other providers work the same way with their own keys.
To add your key:
sk-ant-Your key is encrypted immediately on save and stored securely in your account. It is never visible in the browser or logs.
To get an Anthropic API key:
Anthropic offers a free credit tier for new accounts. Haiku 4.5 is the most cost-effective model for this app - typical note capture costs under $0.001.
MemoRoot supports several LLM providers. Select the company and model in Settings, then enter that provider's API key in the same field. Keys are stored per provider, so switching back to a provider you already set up does not require re-entering its key.
Currently selectable: Anthropic (Claude Haiku 4.5), OpenAI (GPT-5 mini), Z.ai (GLM-5 and GLM-5.3 Flash), and Google (Gemini). Support for Qwen, Moonshot AI (Kimi), and DeepSeek is in development.
Not sure which to pick? See Model pricing and recommendations below.
All prices are set by the providers and charged to your own API key. Listed per 1 million tokens; a typical note capture uses a few thousand tokens, so real cost is a fraction of a cent per note. Prices change over time, so confirm on each provider's pricing page.
| Provider | Model | Input / 1M | Output / 1M | Status |
|---|---|---|---|---|
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | Available (recommended) |
| OpenAI | GPT-5 mini | $0.25 | $2.00 | Available |
| Z.ai | GLM-5 | $1.00 | $3.20 | Available |
| Z.ai | GLM-5.3 Flash | $0.15 | $0.50 | Available (new) |
| Z.ai | GLM-5.2 | $1.40 | $4.40 | Not selectable (underperformed in testing) |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Available | |
| Gemini 3.5 Flash | $1.50 | $9.00 | Not selectable (costs more than any quality gain justifies) |
GLM-5.3 Flash is new. Z.ai's first multimodal GLM-5 model, and by far the cheapest capable option on this list. Through September 9, 2026 a launch promo halves its price to $0.075 input and $0.25 output per 1M; the table shows the normal price it returns to after that. On our hardest multi-source capture it matched GLM-5's note quality while running faster and costing about a seventh as much, so it is available and the results below include it.
How we tested. Every model captured the same deliberately difficult input: a long paste combining several overlapping documents about one subject, with repeated passages and three overlapping biographies. A good result is a single note that keeps every fact, merges the duplicated material into clean sections, and picks a small set of meaningful links. We also measured the real cost and time of each capture. This is a hard worst case; short single-idea captures are easier for every model.
Results.
| Model | Note quality | Measured cost per capture | Capture time |
|---|---|---|---|
| Claude Haiku 4.5 | Very good: clean structure, all facts kept | $0.008 | 15s |
| GLM-5 | Best of the group: cleanest merge and structure | $0.006 | 33s |
| GLM-5.3 Flash | Matched GLM-5 on the hard multi-source merge: every fact kept, conflicting figures attributed per source, one clean note. Clean on simple notes too | $0.0007 | 18s |
| Gemini 3.1 Flash-Lite | Very good after tuning: at its fastest settings it summarized and dropped facts, so MemoRoot runs it with deeper reasoning and a stricter prompt that preserves everything | $0.005 | 18s |
| Gemini 3.5 Flash | Good: near-perfect fact preservation and the fastest accurate run, but left small artifacts and too few links, and costs more than any note quality gain justifies | $0.014 | 12s |
| GPT-5 mini | Good on simple notes; needs its slowest, most expensive setting to handle messy multi-document pastes well | $0.003 to $0.017 depending on settings | 23s or more |
| GLM-5.2 | Poor: left duplicate sections and over-linked | $0.009 | 30s |
Recommendations.
Whichever model you choose, the Usage section in Settings shows your actual spend, token counts, and call counts at the prices above.
We tested both Z.ai models on the same note-capture workload and found that GLM-5.2, despite being the newer and more expensive model, produced clearly worse notes than GLM-5 and was not faster in practice.
GLM-5.2 appears to be tuned for coding and agentic workloads rather than knowledge organization. For structured note-taking, GLM-5 was the better model on every measure, so GLM-5.2 is visible in the model list but not selectable. We will revisit if either model or Z.ai's serving changes.
MemoRoot writes all your notes as markdown files to a storage location you own. MemoRoot also keeps a searchable copy of your note content in its own database to power Search and Ask; your vault remains the source of truth, and deleting a note removes it from both. Pick one storage provider when you connect for the first time:
.md files to a repository you choose. Good fit if you already use GitHub or want version history on every change..md files to a dedicated folder in your Drive. Good fit if you want a familiar, no-setup storage location with no Git knowledge required. Preferred method for MemoRoot / Obsidian dual connected vault.Both providers offer the same MemoRoot features - capture, search, ask, tasks, and Ai enrichment work identically regardless of which one you connect. Pick whichever fits how you already manage files.
You can only have one active vault provider per account. Switching providers later requires disconnecting first - existing notes stay in your original storage and are not moved automatically.
To connect:
Once connected, every note you capture is written as a .md file to your vault repo. You own the data - clone it, back it up, or read it directly on GitHub at any time.
To connect:
Once connected, every note you capture is written as a .md file inside that folder. MemoRoot only reads and writes files within its own folder - it does not access the rest of your Drive.
Google Drive vault access is currently limited to approved test accounts while verification with Google is in progress. If Connect Vault does not work for your Google account yet, use GitHub in the meantime.
Obsidian is exceptional at reading, linking, and deep work, but it has no intake pipeline: raw thoughts need manual structuring before they become useful notes. MemoRoot fills that gap. Capture a messy thought from any device and the AI turns it into a clean, titled, tagged, wikilinked markdown file in your vault, ready in Obsidian moments later.
It runs both directions. Capture on your phone in MemoRoot and a finished note appears in Obsidian, or draft at your desk in Obsidian and it flows into MemoRoot for instant Search and Ask. One plain-markdown vault you own, two tools that cooperate instead of competing, no copy-paste and no lock-in.
Which storage for two-way creation? With a GitHub vault this is fully automatic: notes created by any tool, including brand-new notes you write in Obsidian, are detected and synced. Google Drive is the quickest to set up and is a great fit if you capture in MemoRoot and read or edit in Obsidian, with one limit: on Drive, MemoRoot syncs the notes it created, so a note written from scratch in Obsidian will not appear in MemoRoot (paste it into a capture to bring it in). If creating notes directly in Obsidian is central to your workflow, choose GitHub.
The Drive setup: connect Google Drive as your MemoRoot vault, and point Obsidian at the same folder. This also solves Obsidian's remote vault problem: your vault lives in Drive, reachable from every device, no sync plugins or paid sync service required.
Obsidian works on local files, so the bridge is Google's own desktop app. One-time setup, about five minutes:
MemoRoot folder in your Drive with a notes folder inside, and adds a welcome note.MemoRoot folder and mark it Available offline instead.My Drive/MemoRoot.What you'll see in Obsidian: your notes, sitting right at the top of the vault as plain markdown. MemoRoot keeps its own files (your capture history and a couple of small settings files) in a hidden .memoroot folder; Obsidian always hides dot-folders, so it stays out of your way. One Drive-specific limit to know: MemoRoot syncs the notes it created for you, so a brand-new note you write from scratch in Obsidian will not appear in MemoRoot. To bring one in, paste its text into a MemoRoot capture, or use a GitHub vault, where new notes from any tool are picked up automatically.
From then on, notes you capture in MemoRoot appear in Obsidian after Drive's short sync delay (typically seconds), and your Obsidian edits flow back the other way.
Can't find the MemoRoot folder in Drive? You are probably looking at Drive's Home view, which shows recent and suggested files as a flat list, so folder structure is invisible there. Click My Drive in the left sidebar: the MemoRoot folder is at the top level.
On your phone, skip all of this: MemoRoot itself is the mobile capture surface. Install it to your home screen and capture there; the notes land in the same vault.
Edit freely. When you next open Browse, MemoRoot compares your vault against its search index: changed notes get a small dot and a banner shows how many notes changed, were added, or were removed. Press Sync and MemoRoot re-reads them exactly as you wrote them, with no AI rewriting, so Search and Ask immediately reflect your edits. Renames and deletions of MemoRoot-created notes are followed automatically. The one exception on Google Drive is a brand-new note written from scratch in Obsidian, which MemoRoot cannot see (explained in the storage note above); on a GitHub vault, new notes from any tool are synced too.
This works for GitHub vaults too: edits pushed to your vault repo from any tool are detected and offered for the same one-press Sync.
MemoRoot also protects your outside edits: saving an in-app edit over a note that changed in your vault is refused until you refresh it, and captures that grow an existing note always append to the current file, never replace it.
Obsidian image embeds like ![[photo.png]] pass through MemoRoot untouched. The image file stays in your vault and renders normally in Obsidian; MemoRoot shows a small image badge in its viewer and never moves, renames, or rewrites image files or embed syntax. Image contents are not searchable through MemoRoot, so add a caption in the note body if you need to find one later.
When you capture text, the AI does not store what you typed. It reads your input and produces a structured wiki entry: a title, a filename, sections, tags, an evergreen classification, and concept links. Your words inform the note but the note is the AI's synthesis of them.
This is intentional. Raw text is hard to retrieve and hard to connect to other ideas. A structured note with consistent tags and wikilinks is what makes Search, Ask queries, and backlinks work well over time.
Existing notes grow. When your new input covers the same topic as a note that already exists, MemoRoot handles it according to your Similar content choice (described below) instead of creating a duplicate. The original note is never silently replaced - it always gains, never loses. The post-capture confirmation tells you exactly what happened: a new note created, an existing note updated, or nothing added because the note already covers it.
The Similar content options on the capture form control what happens when a capture covers the same topic as a note you already have. MemoRoot finds candidate notes by meaning, not just by title, so new thoughts on a subject can land in the right note even when you word things differently. Set your preferred default in Settings under Capture Default and change it per capture anytime.
Merge into existing note weaves the new information into the existing note, producing one coherent, well-organized note per topic. Choose this if you want your vault to read like a knowledge base, where each subject builds into a single deep entry over time.
Append to existing note adds only the genuinely new information as its own section at the end of the note. Choose this if you want each contribution to stay visible as a distinct addition.
Always new note never merges. Every capture becomes its own note; only an exact duplicate of an existing note is skipped, and the confirmation names the note that already contains it.
In Merge and Append modes, a capture that contains nothing the existing note does not already cover is skipped with an "Already covered" message rather than padding the note with repetition. Merged and updated notes rise to the top when your notes are sorted by Newest; the note's original creation date is untouched.
The AI watches for signals that a passage should be kept word-for-word rather than synthesized.
Signals it recognises:
"..." or '...')quote:, quote unquote, or quote or unquote written before a passageWhen detected, the passage is placed in a ## Quote or ## Verbatim section using blockquote formatting. Attribution is added on the same line if present (- Author). The rest of the note is structured normally around it.
If the entire input is verbatim - a poem, a full quote - the AI preserves everything exactly and still generates a title, tags, and wikilinks to file it correctly.
LLMs interpret instructions but cannot guarantee perfect behaviour on every input. If a quote is not preserved, try adding quote: explicitly before the passage.
Every note is automatically classified as evergreen or not by the AI at capture time. You do not need to decide.
Evergreen notes contain knowledge that stays accurate over time: concepts, principles, historical facts, scientific findings, ideas. The capital of Michigan will be Lansing next year too.
Not evergreen notes are tied to a specific tool, version, or UI state: where a setting lives in a dashboard, how a particular version of a framework behaves, step-by-step guides for current software. These tend to go stale as products change.
The classification is stored with the note so the vault can later surface non-evergreen notes by age for review, update, or removal.
When checked, the AI adds two supplementary sections to your note beyond what you wrote. Off by default. Set your preferred default in Settings under Capture Defaults.
Context adds factual background from the AI's knowledge that grounds the note: publication history, surrounding events, definitions, relationships to other established concepts. It does not interpret or infer - only established facts.
Analysis adds AI-generated interpretation: what the content implies, likely mechanisms, connections to related ideas, hedged speculation. Uncertain claims are explicitly marked ("likely", "may suggest", "unclear why").
When to use: most useful for creative or literary captures (poems, quotes, historical events) where background knowledge genuinely enriches the note, and for technical troubleshooting notes where understanding the mechanism behind a fix helps future diagnosis.
Knowledge base integrity: enriched sections are saved as ## Context-ai and ## Analysis-ai in the note. The Ask query system recognises these labels and treats them as AI-generated background, not your own knowledge or observations. This keeps your vault trustworthy as a source of your own thinking.
One input does both, and what you do decides which you get:
Search happens as you type. The notes list filters live to notes matching your text - fast, exact, no AI involved. Use it when you know what you are looking for and just want to open the note.
Ask happens when you press Enter (or the ✦ button). Whatever you typed, short or long, you get an answer built from your notes, with the source notes linked below it. Use it when you want an answer rather than a list.
Behind the scenes, Ask picks the cheapest path that can answer well:
The more specific your question, the better the results. A precise question with context gives the system more signal to work with.
Browse lets you explore your vault by tag or title alphabetically, without a query. Use it when you want to orient yourself or follow a thread rather than retrieve something specific. On desktop, the notes list is fully keyboard-navigable: click a title, then move with the up and down arrow keys (each note opens as you go) or type a letter to jump to titles starting with it.
Follow links without losing the answer. Click any note link in an Ask answer to read that note. A "Back to answer" link appears at the top of the note; click it to return to the answer instantly, at no additional cost. You can hop between the answer and its source notes as many times as you like. The answer is kept until you ask something new or pick a different note from the sidebar.
Save an answer. Every finished answer (notes answers and Tasks answers alike) has the same small menu in its top corner. Save as new Note opens a New Note prefilled with the answer text, including wikilinks to the notes it referenced; review and press Capture. Save as new Task creates a task with your question as the title and the answer in the task notes, opened for editing right away.
Copy an answer. The same menu (on both notes answers and Tasks answers) offers Copy markdown, which keeps formatting and note links intact, Copy text, which copies the answer as plain readable text, and Copy rich text, which keeps the formatting for pasting into email, Google Docs or Word.
Asked your notes a question while in Tasks? The Tasks Ask answers from your task list, so it may report no matching tasks. Below every Tasks answer is a link, "Ask your notes this question", which re-runs the same question against your notes without leaving Tasks.
When you open a note, any tag or concept link shown in green exists only in that note. No other note in your vault shares it yet.
Green items are not clickable because following them would just return you to the same note - there is nothing else to navigate to.
As your vault grows and other notes are captured that share the same tag or reference the same concept, the green styling disappears and the item becomes a normal clickable link. This happens automatically the next time you open the note.
Green is a signal, not a problem. A concept that only appears once may be worth capturing more about, or it may simply be specific to that note. Either way, you can see at a glance which ideas are still isolated versus which ones are woven into the rest of your vault.
Over time a vault collects notes that overlap: several captures on the same subject, near-duplicates, the same facts written twice. Settings › Consolidate Notes › Start scans your vault, groups similar notes into topics, and walks you through merging each group into a single note.
For each topic you see the member notes with a similarity score, and you can uncheck any that do not belong. Consolidate selected produces a merged draft for your approval; nothing changes until you approve it. On approval the new note is written and the source notes move to Trash, where they stay restorable for 30 days. Links in other notes that pointed at the old titles are re-pointed to the new note automatically.
Two automatic options exist once you trust the results: Auto this topic commits the current group without an approval stop, and Full auto: entire vault runs every topic start to finish, with Pause and Stop available between topics. Before committing, every draft is checked against the source notes; if details are missing, the model is automatically asked once to rewrite with the missing items named. A draft that still fails the check is skipped and flagged rather than committed, so material never disappears silently.
The recommended workflow for a large vault is auto first, review after: run Full auto to commit the clean merges, then run again with Start: review each topic to finish the flagged ones. There the approval screen lists exactly which details are missing from a draft, so you decide whether it matters and approve, or discard and redraft.
You choose which of your configured AI models runs the job; the choice applies to that run only and does not change the model used for capture. GLM-5 is the recommended default: consolidation is faithful rewriting, not creative work, so a bigger model rarely improves it. Reach for a stronger model on topics that combine many long notes or dense technical detail, and feel free to use your cheapest model on small groups of near-duplicates. Each run shows its total AI cost when it finishes.
The filter pills at the top of the Tasks tab control which tasks are shown:
The Priority sort groups high and medium priority tasks above the rest. The Age sort cycles through Newest, Oldest, and off. In Today and Scheduled views, Age sorts by due date rather than creation date. Both sorts can be combined.
Set priority on any task via â‹® menu › Set priority. It cycles through None, Medium (amber title), and High (red title). The menu label updates to show the next step so you always know what one tap will do.
Set a due date via â‹® › Set due date, or click an existing date badge directly. Overdue tasks show their date in red. In the Today and Scheduled filters, the Age sort orders by due date.
Add a note to any task via â‹® › Add note. A note badge appears on the task row - click it to show or hide the note inline. Edit the note directly in the text area that opens; it saves when you click away. Any web links in the note appear as clickable links just below it, and open in a new tab.
Notes are also visible and editable in the expanded view.
Open any task in full-screen via â‹® menu › Expand. The expanded view shows the task title and notes in a large editable area.
Edit the title or notes directly. Changes save automatically when you collapse - via â‹® menu › Collapse or the Escape key. Nothing is lost if you navigate away; the save happens the moment you close the view.
Due date and priority are shown as badges below the title and can be changed from the kebab menu inside the expanded view.
Type a question in the task input and click Ask to get an AI answer about your tasks.
The AI reads your full task list and answers directly: what is overdue, what is high priority, what is coming up this week. Relevant tasks appear below the answer as fully interactive rows - you can complete them, set a due date, change priority, or expand them without leaving the view.
After the answer appears, the Ask button resets. To return to the normal task list, click Ask with the input empty, or just start typing a new task and press Enter.
Set an alarm on any task via â‹® menu › Remind me. Choose a preset (In 10 minutes, In 1 hour, This evening, Tomorrow 9:00) or Pick time for an exact date and time. At that moment a notification arrives on every device where you have allowed notifications for MemoRoot, even with the app closed. A gold clock badge on the task row shows the reminder time; click it to change or clear it.
The first time you set a reminder, your browser asks for notification permission. Permission is per device: allow it on each device where you want reminders to arrive. Settings › Tasks lists your registered devices and has a Send test notification button to verify delivery.
When a reminder fires with the app open, a card appears with OK, snooze options (5 min, 10 min, 1 hr, Tonight 9 pm, Tomorrow 9 am), and Open task. With the app closed, you get a system notification instead, with a Snooze 5 min button on phones and two snooze presets on desktop; tapping its body opens the app on the Tasks view with that task highlighted. Reminders are silent banners by default: turn on Reminder sound in Settings for your device's standard notification sound plus a soft in-app chime.
If a reminder fired over an hour ago and was never dismissed, or could not be delivered, a missed reminders banner appears at the top of Tasks. Expand it to snooze each one, dismiss it with OK, or jump to the task.
Platform notes: on desktop the browser must be running for notifications to arrive. On iPhone and iPad, add MemoRoot to your home screen first (iOS 16.4 or later); the app icon then shows a number badge for unseen reminders, cleared when you open the app. On Android, icon badges are drawn by your launcher from the notification tray and are sometimes off by default: long-press the MemoRoot icon › App info › Notifications and turn on app icon badges. Reminders are separate from due dates: a due date drives the filters, a reminder is an alarm, and a task can have either or both.
When you save an API key, it is encrypted immediately on the server using AES-256-GCM before being stored. The encrypted value is kept in MemoRoot's database, keyed to your account and only accessible via the server-side API - it is never exposed to the browser, to client-side JavaScript, or in any dashboard view.
At request time, the key is decrypted in memory inside the edge function, used to make the API call, and then discarded. It is never logged, never returned in a response, and never written anywhere in plaintext.
The encryption key itself is a 32-byte secret stored as a Cloudflare Workers environment variable, not in the codebase or any database. Without it, the stored ciphertext cannot be decrypted.
Save your API key in a password manager. If you lose access to it, you can always generate a new key from console.anthropic.com and re-enter it in Settings.
MemoRoot uses Clerk for the full authentication layer: sign-in, session management, and per-user data storage. This means we never store or handle passwords directly - Clerk manages credential verification and issues signed JWTs that our server validates on every request.
Vault access is provider-specific: the GitHub connection uses a GitHub App installation, while the Google Drive connection uses an OAuth token pair (access and refresh) that auto-renews. Both are retrieved server-side on each request, so neither is ever held in browser storage.
User preferences (selected model, company) and your encrypted API key are stored server-side in MemoRoot's database, scoped to your account. Neither has a client-side access path.
The complete picture, so nothing here is a surprise: