Memory (Beta)
Last reviewed: 2026-11-19
The memory holds what the AI should know about you and your work without being told again: what you do, how you like your answers, who your clients are. It is two files in your vault. Nothing gets into it without your yes, and you can read, change and delete every entry.
Two places
Always included goes into every new conversation. Keep it short: it has room for 2,000 characters, and a bar shows how full it is. An entry that no longer fits is marked No room left — not included — it is saved, but not sent along. Entries are included in the order they stand in, so what matters most belongs at the top; you change the order in the file.
On demand is not sent along. When a question may depend on something you told the AI earlier, it looks there, and the conversation shows Looking in the memory. This is the place for what matters only sometimes: one client’s terms, a decision and its reason.
Opening the memory
On the desktop, choose Memory in the AI tab, next to Skills. On the phone it is Memory on the AI screen, or Skills and memory, the last entry of the Skills menu of an empty conversation. Settings → AI in this vault leads there as well: Open memory under Skills and memory. The file names of the memory stand behind the ⓘ of the card Always included; the rules of the vault (AGENTS.md) are reviewed and approved right there in the memory view.
Adding an entry yourself
New entry asks for three things: the text — one thing per entry, a single line of at most 500 characters —, the place (Always included or On demand), and whether it is for Only models on this device. Tick that for anything no cloud model should be told.
The ⋯ button of an entry — on the phone, a tap on the entry — offers Edit, moving it to the other place (Always include or Only on demand) and Delete.
Letting the AI remember something
Say it in a conversation: “Remember that I bill per day, not per hour.” The AI drafts an entry; it never writes one itself. Under its answer a card Draft · Memory entry shows the whole text. Choose the place, then Remember — or Discard. A draft you have not decided on waits under To decide, and the memory says how many are waiting.
“Forget that …” works the same way: the card reads Draft · Remove from the memory, and Remove takes the entry out. When you tell the AI that something has changed, the card shows under Replaces which entry the new wording takes the place of.
A finished conversation can suggest entries too: Learn from this conversation reads it once more and leaves drafts, each with its evidence. How that works, and what such a review may suggest at all, is described in Skills.
A rule is not a memory
“Always answer in German” is nothing to know — it is something to do. A rule like that does not go into the memory: it becomes a line of the Instructions of the vault (AGENTS.md), which every model gets as an instruction. Add one with Add a rule under Rules for the AI, or ask the AI; its card then reads Draft · Rule for the AI, with the button Add as a rule.
Like all instructions, the file has to be approved on each device before it counts there (see Skills). A rule you add on a device where the file was already approved counts there at once; your other devices ask you first.
Tidying up
A memory that has grown repeats itself. Under Tidy up, the memory view names what this device noticed by itself, without asking a model: two entries that say almost the same, an entry that is more than a year old, and entries of Always included that no longer fit. Compare shows the two entries together, each with everything its menu can do.
For a closer look there is a skill. Memory care reads the entries with the model of a conversation and drafts what belongs together and what can go; the row Have the memory looked through starts it — it stands under Tidy up whenever something is named there. What it suggests are drafts like any other: nothing changes before you accept it, and where two entries contradict each other it asks instead of deciding.
Writing style
A draft of the AI sounds like the AI. The section Writing style describes how you write, so that a draft can sound like you — measured from notes you choose yourself.
Choose examples… opens the note picker of Pin a note…; choose up to eight notes you wrote yourself. This device measures in them, without a model and without sending anything: how long your sentences and paragraphs are, lists or running text, headings, how you address people (in German “du” or “Sie”, in French “tu” or “vous” — where a language shows it in its words), bold, emojis, exclamation marks and dashes. Properties, code and quotes do not count. Every feature stands there with the numbers it rests on, and its ✕ takes it out on its own; Bring back brings back what you took out. No word and no phrase of your notes is kept: the device keeps the paths of the examples and what you took out, and measures again whenever the style is shown or used.
The style is used only where you choose it: My style, the button beside the globe under the message field, gives it to the conversation you start next — like the internet, fixed from its first message. The conversation then says With your writing style, the send overview has a row Writing style, and a draft from such a conversation says that it was drafted with your writing style; it remains the AI’s draft, signed with its model. The AI follows the style in what you keep for yourself — a note, a journal entry, a rewrite of your text —, not on its own in an e-mail, an invitation or a comment; ask for it there, and it does.
The style lies on this device, not in the vault: on another device you choose its examples there. An example your privacy rules keep from a model gives that model none of its features — the send overview counts it under Kept back —, and what a model on this device writes with features of such a note keeps the note’s rules.
Privacy
- An entry can carry a rule of its own: Not to cloud models, Not in conversations with the internet. A model on this device gets every entry that is readable and short enough — except in a conversation with the internet, where an entry marked Not in conversations with the internet stays out here too.
- An entry the AI drafted in a conversation that read notes under a privacy rule gets the same rules — the card says The entry gets the privacy rules of the notes this conversation rested on. What came from a note that has to stay on this device does not reach a cloud by way of the memory.
- Your privacy rules hold for the two files as well: a folder rule for
.agent/keeps the whole memory from the cloud. - The send overview has a row Memory — how many entries go along — and counts the entries your rules keep back under Kept back. It never names an entry.
- To the AI an entry is information, not an instruction: a sentence in the memory gives it no rights.
- A conversation keeps what the memory gave it at its start — the entries under Always included. An entry you delete goes to no new conversation; conversations that have already begun keep what they were given, and what they look up later comes from the files as they are then.
Switching it off
Use the memory on this device is on until you switch it off. Off, a conversation on this device gets nothing from the memory and adds nothing to it. The files stay as they are, and every device decides for itself.
The two files
.agent/active_memory.md (always included) and .agent/MEMORY.md (on demand) are plain Markdown. Every entry is a list item, and headings group entries. What Plainva knows about an entry stands in a comment behind it:
## Clients
- Harbour Studio pays within 14 days.
- I bill per day, not per hour. <!-- plainva: added=2026-10-09; by=assistant; source=Offer for Harbour Studio; deny=cloud -->
added and by say when the entry was added and whether you wrote it or accepted a draft, source names the conversation a draft came from, and deny holds its rules (cloud, web). You can edit the files in any editor. In Plainva, Open file opens either of them.
Plainva does not guess. An entry whose comment is damaged is marked Rules unreadable — goes to no model until you repair the comment or add the entry again. Text that is hidden in a comment or in invisible characters is never sent; the entry then shows how many hidden parts were left out. Of a file at most 2,000 entries are read, and a file larger than 256 KB is not read at all.