Give Claude a memory you can read and correct
Claude in the app and Claude Code at your desk each keep their own memory, and neither reads the other's. A bullet in your own outline is a memory both can reach, and one you can read and fix yourself.
Claude in the app and Claude Code at your desk each keep their own memory, and neither reads the other's. A bullet in your own outline is a memory both can reach, and one you can read and fix yourself.
When Claude reaches a step only you can take, it should not ask in a chat that scrolls away. It should write a checkbox task into a list you both read, and carry on with the rest.
Claude on your phone reaches MCP servers from Anthropic's cloud, so a note app whose server runs on your laptop is out of reach. Of 21 note apps, 9 host a server Claude can use from the phone; 12 do not.
Your notes hold the project you want help with and a lot you never meant to show an agent. A sentence in the prompt asks Claude to stay out; a grant on the server makes it unable to look.
Two Claude Code sessions in one repository can share a CLAUDE.md. Once one of the agents is the Claude app or ChatGPT, the decisions both need have to live somewhere both can read, and you should be able to read it too.
A list of work that your agent reads before anything else keeps requests from dying with the session. I kept one for two weeks, then let the agent write it from what I say in the chat. What each version is good for.
Remote Control puts a Claude Code session from your own computer on your phone. What it does not change is that a session is a long transcript. How I steer the agents that build Pando from the phone with two short lists and checks that a program runs, not the agent.
An MCP agent that reads your whole outline at the start of every session pays for it each time. Asking only for what changed cost 642 bytes after five edits and 100,373 after a busy week, against 106,606 for one cheap re-read.
An agent that worked for an hour writes a report you will not finish on a phone. Written as nested bullets, the top layer is five findings with their numbers, and the evidence is one tap under each, still there the next day.
When an agent keeps facts that repeat, such as quotes, statuses or dates, a table in your outline makes each update one cell and shows you the gaps at a glance on your phone, and it turns back into a list with one change.
The case for giving agents command-line tools instead of MCP servers is strong where the agent has a terminal. Claude's chat on a phone and ChatGPT in a browser tab have none, so a tool that ships only a CLI does not reach them. Which agent can run what, with vendor pages read on 14 November 2026.
If Claude Code runs on the laptop where your vault is, you need no MCP server. The other common cases, from the phone app to ChatGPT, each need something different switched on.
Forward a mail to one bullet's own address and it arrives in your outline, subject as the line, message as the note, where Claude reads it without access to your mailbox. Forwarded text is still untrusted input, and the post says what Pando does about that and what it does not.
Claude merkt sich bei Free, Pro, Max und Ultra von sich aus, was Sie im Chat erwähnen, und löschen Sie den Chat, bleibt die Erinnerung stehen. Was Anthropic, OpenAI und Pando in ihren Datenschutzerklärungen und Hilfeseiten zu Speicherort, Drittlandübermittlung, Löschung und Export schreiben, Stand 14. November 2026.
Bevor Sie Claude oder ChatGPT über einen gehosteten MCP-Server an Ihre Notizen lassen, lohnen sich sechs Fragen an dessen Betreiber, vom Serverstandort und dem Auftragsverarbeitungsvertrag bis zum Widerruf des Tokens. Hier stehen sie mit Pandos Antworten, auch dort, wo Pando schlecht abschneidet, und mit einem Test, den Sie selbst wiederholen können.
GTD usually fails between the apps: the capture is in one place, the projects in another, and the weekly review means opening all of them. In one tree the five lists sit on one screen, and Claude drafts the filing for you to check.
A weekly review slides when it costs a walk through five lists and leaves nothing behind. A master checklist with dated runs in one tree makes it shorter and gives it a record, and Claude can prepare each run.
Merlin Mann gave every item in an inbox five ways to leave. They work on any inbox, a mail folder, a notes app or the list Claude writes into, as long as the places they lead to are close by.
Do it now if it takes less than two minutes. David Allen wrote that for the moment you empty an inbox; applied to a morning of focused work, it can eat the morning.
Things other people owe you are in nobody's task list, so you notice them when the consequence arrives. A Waiting For list with who, what and the dates you asked is a short list to keep, and Claude can tell you what has gone silent.
An idea you are not doing this year still sits on your project list and makes the list harder to trust. Parking it on a someday list with two dates is a smaller decision than deleting it, and one you can make today.
Doing the hardest task first is good advice that tends to fail at 08:45, because that is when you have to pick it. Brian Tracy's own book also says to plan the day the evening before, and that half moves the choice out of the morning.
A to-do list holds eleven tasks for a day with four free hours and gives no warning. Giving each task a block of the day shows in the morning what does not fit, while there is still time to choose.
Urgent work announces itself, so three quadrants of the Eisenhower matrix fill up on their own. The fourth, important and not urgent, stays empty unless you put things there and give them a day.
A goal for the whole year leaves room to postpone until autumn. Twelve weeks with the week number in the title leaves little room, and a Friday score tells you whether the plan was too big.
Personal OKRs usually fail at the key results, which turn out to be smaller wishes. The repair is one rule, a number with a source, and a short check each week.
When a project has stalled for months, check whether anyone wrote down where it ends. Five lines under the project, written in a few minutes, say what done means, what is out, who decides, the first step and the cost of dropping it.
A note about a supplier fits under four topics, so it ends up in a different one each time. PARA asks how soon you will act on the note instead, which is easier to answer.
Luhmann's box connected its notes in three ways, and most digital copies keep only the box. An outline can give each note a place behind the one it continues, links to others and a spot under every topic it belongs to.
Exported highlights pile up unopened, because each one is somebody else's sentence. Write the idea in your words as the parent, hang the quote under it, and add where it applies to you.
A meeting note is a note about a Wednesday, and nobody reads those again. Write four kinds of line, then move three of them to the places where the work happens.
Experienced people skip steps they know well, under pressure. Gawande's checklist rules, with clean masters you never tick and dated runs you do, as a tree you can copy.
A streak counter shows one number and resets it on the first missed day. A table with a column per day keeps the blanks, so you can see which days you miss and change the habit to fit them.
Personal Kanban has two rules, a board and a limit on how many things are in progress at once. The board is the easy part; the method relies on the limit, and a lane for parked work keeps the limit workable.
Four lines, written the week a project ends: what you expected, what happened, why they differ, and one rule for next time as a checkbox. The rule is the line you act on.
Notion's hosted MCP server reaches every page your account can reach, and I found no read-only or one-page setting for it. Pando's server can be read-only or held to one bullet, and the agent's memory is a bullet you can read. Checked on 14 November 2026.
Workflowy's MCP server runs inside its desktop app, so Claude on your phone or on claude.ai has no official way into your outline. Pando's server is hosted. What else differs, and where Workflowy is still the better outliner, checked on 14 November 2026.
Roam asks $165 a year and has no free plan. Pando is free up to 5,000 bullets and 3 EUR a month past that, but it has no block references, no queries and no graph view. Who should move, who should keep paying, and how a Roam page arrives, checked on 14 November 2026.
Logseq is now two apps. Logseq OG keeps your Markdown files and gets no new features, and the new Logseq is a database in beta. If files on your disk are why you use Logseq, stay on OG; if you are choosing anyway, here is what Pando offers instead, checked on 14 November 2026.
Tana Outliner has supertags and its own AI, and its hosted MCP server is a beta on paid plans. Pando has neither supertags nor built-in AI, and its hosted server is on the free plan. Checked on 14 November 2026.
Claude on your phone has no official way into a Dynalist outline; the MCP servers I found run on your own computer. What Dynalist still does better, what Pando adds for an agent, and how Dynalist dates arrive in Pando, checked on 14 November 2026.
Moving into Pando takes one file if your old outliner exports OPML, and one import per page if it exports Markdown. What each kind of file carries, what never comes across, and a short script for a folder of pages, tried on 14 November 2026.
Für eine Antwort schickt Notion AI passende Seiten aus Ihrem Workspace an Sprachmodelle. Notion nennt acht Auftragsverarbeiter mit Bezug zu Sprachmodellen, aber nicht, welcher Ihre Seiten im Einzelfall bekommt. Pando hat keine eingebaute KI, dafür weder EU-Hosting noch AVV. Ein Vergleich nach den Seiten beider Anbieter, Stand 14. November 2026.
Workflowys eigener MCP-Server läuft nur in der Desktop-App, die dabei offen bleiben muss. Pando hat seinen MCP-Server im Netz, kommt von einer Firma in Soltau und hat deutsche Rechtstexte; die App ist bei beiden englisch. Ein Vergleich nach Workflowys eigenen Seiten, Stand 14. November 2026.