Memory

Pia’s Memory is the personal knowledge it builds up about you — your preferences, the people and projects you mention, important dates — and draws on to give you answers that fit you better.

Everything is stored in a vault: a folder of plain Markdown files on your computer. Pia owns and writes these files, but they’re just text — you can read, edit, or open them yourself at any time.

Memory, vault, objects — the words Pia uses

Pia uses three words for closely related things, and Settings, the Memory view and this guide each lean on a different one. They fit together like this:

The vault is the folder on disk. It always lives in a Vault subfolder of your Assistant files folder, and it holds two things: a memory folder that Pia writes, and a sources folder for documents you add.
Memory is what’s inside the vault’s memory folder, and it’s what the Memory view in the sidebar shows you.
An object is one entry of memory — one thing Pia knows. The header count “48 objects · 52,6 KB” is telling you that Pia has 48 separate entries taking 52.6 KB of disk.

So “the vault” is the container, “memory” is the contents, and an “object” is one item. Settings names the Assistant files folder and shows Memory vault: …\Vault underneath it because the vault always sits inside that folder — move one and the other moves with it.

Vault size, and when to prune

A count like 48 objects is not a lot, and there is no pruning you are expected to do. Memory objects are small Markdown entries — a few dozen of them is a normal working vault after a bit of use, and the whole thing in that example is 52.6 KB, which is less than a photo. There is no size limit to stay under and no performance cliff to avoid, so there’s no housekeeping you’re expected to do.

Delete a memory when it is wrong or out of date, not to keep a number down. A stale fact Pia keeps repeating is worth removing; a large but accurate vault is exactly what the feature is for.

How memory works

Pia remembers automatically. While you chat, it quietly records useful things. New memories appear in the Memory view on their own — you don’t have to save anything.
You stay in control. Select a memory to read it, change its text, or delete it so Pia stops using it.
Search understands meaning, not just exact words — describe what you’re looking for.
It’s all local. Your memories are files in the vault’s memory folder on this computer.

Browsing your memories

Open Memory from the sidebar. The left column lists your memories, grouped into collapsible categories:

Personal Profile, Contacts, Preferences, Notes, Projects
Topics — knowledge pages, further split into People, Organizations, Products, Concepts, Regulations, Technology, and Other

Each group header shows a count and when it was last updated; items inside are sorted by title. If you haven’t built up any memory yet, you’ll see “No memories yet — Chat with Pia to start building memory.”

The toolbar at the top of the Memory view has five buttons:

Button	What it does
Back	Returns to the page you were on before following a `[[link]]`, like a browser’s back button.
Home	Returns to the top of the Memory view, clearing the selection so Vault at a glance shows again.
Refresh	Re-reads the vault from disk. Use it after editing files in Explorer.
Open memory vault	Opens the vault folder in Windows Explorer.
Help	Opens the “How memory works” dialog.

Where the categories come from

The categories down the left — Personal Profile, Contacts, Preferences, Notes, Projects, and the Topics group with People, Organizations, Products, Concepts, Regulations, Technology and Other — are fixed. They are built into Pia, and you can’t add, rename, remove or reorder them.

They aren’t folders you file things into either. Each memory carries a type, and the category headings are just Pia grouping entries of the same type together. Categories with nothing in them don’t appear at all, which is why your list may show fewer than the full set.

The PROFILE, PREFERENCE and NOTE tags on entries

The small tag on a memory entry is its type — what kind of thing Pia thinks it recorded:

Tag	What it means
PROFILE	A fact about you, or a contact.
PREFERENCE	Something you like, want, or want avoided.
NOTE	General knowledge that isn’t one of the above. This is the fallback.

Pia sets the type when it writes the memory, and it decides the category the entry appears under. The type isn’t editable from the Memory view — there is no dropdown to change PROFILE into PREFERENCE. What you can edit is the memory’s text.

If the type is genuinely wrong and it bothers you, the two practical routes are to delete the entry and tell Pia the fact again more plainly, or to open the vault folder and edit the entry’s type field in the Markdown file directly, then hit Refresh.

Adding a memory yourself

You can’t add a memory directly from the Memory view — it has no “New memory” button. Memory is something Pia writes as you talk, so the way to add one is either to tell Pia in a chat (“remember that I prefer metric units”), or to create the Markdown file yourself in the vault’s memory folder and press Refresh.

What you can do from the view is read, edit and delete any entry that already exists — see the sections below.

Vault at a glance

When nothing is selected, the right-hand panel shows Vault at a glance — a Composition by category bar that visualizes how your memory breaks down, with a colour-coded legend and counts.

Below it is the Source documents list — your own files that Pia has compiled into knowledge. See Documents & Auto-Ingest for how to add them.

Reading the composition bar

The Composition by category bar is descriptive, not a health check, and it asks nothing of you. Seeing Products 35% and Technology 29% means your vault currently holds more entries about products and technology than about anything else — usually just a reflection of what you’ve been talking about or which documents you added.

There is no target balance, no category that should be bigger, and nothing to rebalance. The bar is there so you can see at a glance what Pia knows a lot about, and spot a surprise — a category you never expected to be large is a hint that something got recorded that you might want to read or delete.

What “Compiled into 31 topic page(s)” means

A topic page is a Markdown page Pia writes about one subject — a person, a company, a product, a concept — pulling together everything your source documents say about it. One document usually mentions many subjects, so a single file can produce many topic pages; “Compiled into 31 topic page(s)” means Pia found 31 notable subjects in that document and wrote or updated a page for each.

Topic pages are ordinary memory, so yes, you can read them. They appear in the left column under Topics, grouped by kind (People, Organizations, Products, and so on). Click one to read it like any other memory, follow its [[links]] to related pages, and edit or delete it if you want.

Topic pages live in the vault at memory/topics/. Your original documents in sources/ are never changed — Pia only reads them. See Documents & Auto-Ingest.

Topic pages are written in the language Pia is set to, whatever language the documents behind them are in. A vault fed by mixed sources reads as one language rather than switching page by page.

Reading a memory

Click a memory to open it in the inspector. You’ll see its type, title, and when it was last updated, with the content rendered as formatted Markdown.

Links between pages — [[double-bracket]] links are clickable; use Back and Home to navigate between pages the way you would in a browser.
Copy — copy the memory’s raw text to the clipboard.

Editing a memory

Click Edit to open the memory’s Markdown in a text editor, make your changes, and click Save. Pia updates just that memory and leaves the rest of the file untouched.

To remove a memory, click Delete and confirm. Pia will no longer use it.

Searching

Use the search bar above the list to find memories. If an embedding model is available, Pia uses hybrid search (meaning + keywords) for more accurate results; otherwise it falls back to keyword matching. The status bar at the bottom shows whether the embedding model is Ready and which model is in use.

The embedding model

The embedding model is what lets Pia search your memory by meaning rather than by exact words — so that asking for “the place I get my coffee” can find a memory that only ever said “café”. It turns each memory into a list of numbers (an embedding, 384 of them per entry) and compares your question against those numbers.

The embedding model is a small multilingual model that runs on your own machine. Pia downloads it once, keeps it in your local app data, and never sends your memories anywhere to use it. It is not an API model and it costs nothing to run — which is also why it works with no AI provider configured and with no internet connection.

The status line at the bottom of the Memory view shows whether it is Ready and its vector size. If it isn’t downloaded yet, search quietly falls back to keyword matching — you’ll still find things, just by the words you actually wrote.

Regenerate Embeddings

Regenerate Embeddings rebuilds the meaning-index for every memory in your vault from scratch. It asks you to confirm first (“This will regenerate vector embeddings for all memories. This may take a moment.”), then works through your entries with a progress indicator.

How long it takes scales with how many memories you have. For a normal vault of a few dozen entries it’s seconds; a large vault built from many ingested documents can take a few minutes. It runs locally, so it costs no tokens and no money.

You rarely need it. Pia embeds each memory as it’s written, so the index normally stays current on its own. Reach for Regenerate Embeddings when:

You downloaded the embedding model after already building up memory, so older entries were never embedded.
You edited memory files directly in Explorer and search isn’t finding the new wording.
Search results feel wrong or stale in a way that Refresh doesn’t fix.

If the model hasn’t been downloaded yet, the button tells you so (“Please download the embedding model first before regenerating embeddings.”) rather than doing nothing.

If the embedding model changes

Embeddings from one model can’t be compared with embeddings from another, so if Pia ever moves to a different embedding model, the numbers stored against your existing memories stop matching.

Your memories themselves are never at risk. They are plain Markdown files — the embedding index is a derived search aid sitting alongside them, not the data. The worst case is that meaning-based search gets less accurate until the index is rebuilt, which is exactly what Regenerate Embeddings is for. Nothing is lost, and no text is re-sent anywhere, because the whole process happens on your device.

@ Commands

In the Assistant chat, type @ to point Pia at one specific thing. An autocomplete popup appears as you type: pick a domain first, then the individual item.

Syntax	Example	Description
`@Memory`	`@Memory`	Point the assistant at your memory and topic pages
`@Memory:Name`	`@Memory:ProjectAlpha`	Reference a specific memory
`@Memory:"Multi Word"`	`@Memory:"My Project"`	Quote titles with spaces

There are five domains:

Domain	What it tags
`@Files`	A file in your assistant folder — top of the list and picked by default, and only offered while file access is on
`@Memory`	Any page or record in your vault. The picker lists the whole vault, filtered as you type, not just your Personal Profile
`@Todo`	A task on your Kanban board
`@Reminder`	A reminder
`@Routine`	A routine — asking for it to be run starts it

Where your memories live

The vault lives at <your assistant files folder>\Vault and contains two folders:

memory/ — the memories you see in the Memory view.
sources/ — your own documents that Pia reads for reference (see Documents & Auto-Ingest).

Click Open memory vault in the Memory view to open this folder. You can move it to a different location — see General Settings → Assistant files and memory folder.

What leaves your device

Opening the Memory view, browsing, reading an entry and searching are all local. Nothing is sent to an AI provider just because you looked at your memory, and meaning-based search runs on the embedding model on your own machine.

Your memory does reach an AI provider in three situations, all of them things you asked for:

When it’s relevant to a chat. Pia recalls matching memories and includes them in the message it sends, so the assistant can use what it knows about you. Tagging a message with @Memory does the same thing deliberately.
When Pia writes a memory. Deciding what’s worth remembering happens as part of an ordinary chat turn, so the conversation it came from was already being sent.
When a source document is compiled. Turning your own documents into topic pages sends their contents to your configured provider. See Documents & Auto-Ingest.

If PII Tokenization is switched on under Settings → General → Privacy, personal details in recalled memory are replaced with tokens like [Person_1] before the request leaves your device, and resolved back afterwards. See the Privacy guide.

Storage and sync

Your memory vault is stored as plain Markdown files on this device, and it keeps working with no account at all.

When Cloud Sync is on, your memory syncs across your devices along with your conversations and todos, end-to-end encrypted before upload on plans that include E2EE.

The encryption setup banner stops that sync, and nothing else. While the orange “End-to-end encryption setup required to sync your data” bar is showing, this device holds the key it needs to sync but hasn’t been set up yet — so memory you build here stays here, and memory from your other devices doesn’t arrive. Everything in the Memory view still works normally: Pia keeps recording, searching, recalling in chat and compiling documents, all locally. Finish the setup and both directions catch up.