Your memory
Everything palaia stores is a plain text file on your machine — nothing lives only inside a database you’d need special software to read. This page explains what’s actually happening, in the order things happen.
Notes are files
Section titled “Notes are files”A note is a short document: a title, a body, a couple of tags, and a folder it lives in. Open one in any text editor and you’ll recognize it instantly — it reads like a well-kept wiki page, not a row in a spreadsheet. Every change to every note is tracked, the same way source code is tracked, so you can see exactly what changed and when, and nothing is ever silently overwritten.
Notes can point at each other — “see also” links inside the text — so related ideas stay connected without needing a folder structure to capture every possible relationship. The dashboard’s Memories page shows a note alongside the handful of others it’s directly connected to, so you can follow a thread without getting lost in everything else that’s ever been saved.
Open them in your own editor, today
Section titled “Open them in your own editor, today”Because a memory is just a folder of Markdown files, you can point a
note-taking app straight at it — Obsidian opens the folder as a working
library with no export, no conversion and no plugin, and the links between
notes light up in its graph the way they would in any other library. Any
other Markdown editor works the same way; so does a plain text editor, and
so does grep.
Two things worth knowing if you do:
- Edit freely. palaia watches the folder, so a note you change in Obsidian is picked up and searchable within moments — you are not editing a copy.
- It is not a second backup. Reading the folder from another app is convenient, not protective: the same disk failure takes both. Use Back up & restore for that.
The folder lives wherever your hub keeps its data (on the usual setups, a
vaults/ directory inside it); the dashboard shows the exact path for each
memory.
For the full setup — syncing with git when the hub runs on its own box (the usual case), or pointing Obsidian straight at the folder when palaia runs on your own machine — see Edit in Obsidian.
How something you say becomes a note
Section titled “How something you say becomes a note”When a connected AI tool saves something on your behalf — a decision, a correction, a detail worth keeping — it doesn’t write directly into a polished note. It drops a short, structured record: what it concerns, why it’s worth keeping, the actual content, and where it came from. Think of this as a stack of quick, unreviewed entries rather than a finished document.
A quiet background process reads that stack shortly after and does one of two things:
- Files it away on its own, when the job is simple: create a new note, or add a new detail to a note that already exists. This happens without asking you anything.
- Proposes a change instead of making it, whenever the job would touch something that already exists in a bigger way — merging two notes, renaming one, retiring one that’s gone stale. Nothing like that ever happens automatically. You see the proposal and approve or reject it yourself — in the dashboard under Memories › Needs review, or in the review view your AI tool can show you.
That split is deliberate: adding something new is low-risk and constant, so it stays automatic; changing something that already exists can lose information if it goes wrong, so a person stays in the loop for every instance of it.
When a new note looks like an old one
Section titled “When a new note looks like an old one”Knowledge goes stale quietly: an AI tool saves “the health check returns 404” while an older note still says “returns 200”, and from then on both answers come back when anyone asks. To catch that moment, palaia compares every new note with what is already in your memory. When one means nearly the same thing, the AI tool gets a short heads-up along with the “saved” confirmation, naming the existing note — so it can update that note instead of keeping two versions side by side.
The new note is saved either way; the heads-up is advice, not a gate. It
only works once meaning-based search is running (see below), and it can’t
tell an update from a note that is merely written the same way, like two
weekly meeting notes — which is why it says possible overlap. How close
counts as “nearly the same”, or switching the check off, is set in
config.yaml on the machine palaia runs on:
nudges: similar_note_check: true similar_note_threshold: 0.7 # 0.5-1.0; higher = fewer, surer warningsWhen a note has a known problem
Section titled “When a note has a known problem”Some problems with a note don’t show in its text. A link can still use the old name of a note that was renamed since, two notes can claim the same address, or a file saved in an old text encoding can’t be edited through palaia. The hub checks your memory for these once each time it starts. When an AI tool then reads one of those notes, it gets a one-line heads-up with what to do next: fix the link (and on which line), or put the change in a new note instead of editing.
Before the heads-up is shown, palaia checks it against your memory as it is
now — a link someone repaired since, or a duplicate address that was
resolved, is not mentioned — and as soon as a note changes, palaia forgets
what it knew about it. Problems that come up after the hub started show up
after its next restart. For the complete check, run
palaia doctor (see Troubleshooting).
Finding things again
Section titled “Finding things again”The search bar in the dashboard looks for both the words you type and notes that mean roughly the same thing even if the wording differs — so “deploy process” can surface a note titled “release checklist” if that’s what actually answers the question. The same lookup is what an AI tool uses when it checks your memory before answering something on your behalf; see Your first shared memory for what that looks like from the AI’s side.
When an AI tool looks something up and several notes answer about equally well, the ones your tools keep pulling up come first — and a note nobody has asked for in months slowly gives that place back. Nothing is deleted or moved for this: it only nudges the order of close matches, never whether a note can be found at all.
Is your AI tool actually using it?
Section titled “Is your AI tool actually using it?”The dashboard’s AI apps page shows, for each AI app, how often it looked something up and how often it saved something over the last seven days. It also says how many of its sessions saved before looking anything up, which is the sign of a tool that writes without reading what is already there. Only counts are kept, never what was searched for or saved, and they survive a restart of the hub.
An AI tool that is unsure whether it can reach your memory can ask the hub directly: every connection has a status check that lists the memories it may read and write.
More than one memory
Section titled “More than one memory”You can keep separate memories — work and personal is the common split — and decide per AI tool, or per connection, which one it can see. They stay genuinely separate: nothing crosses from one into the other on its own. Profiles & access covers how that’s set up.
A temporary memory for one task
Section titled “A temporary memory for one task”Some work needs a lot of reading that is only useful for a short while: research for one trip, the sources behind one report, the background for one migration. Putting all of that into your everyday memory would bury the things you actually want to keep. Instead, give the task a memory of its own — a temporary one:
- Create it, optionally with an end date. Your AI tools see it as a separate memory with its own name, so their searches in it stay on the task and never mix with your everyday notes.
- Work in it. Ask your AI tool to collect the sources and notes for the task there, and to answer from it.
- Keep what matters. When the task is done, pick the few notes worth keeping and copy them into your everyday memory. They arrive with their titles, tags and links intact; if a note with the same name is already there, nothing is copied and you’re told which one clashes.
- Close it. Closing takes the temporary memory away from every AI tool and moves its folder into an archive next to your hub’s data. Nothing is deleted — you can still open the archived folder, or delete it yourself when you’re sure.
Nothing happens on its own when the end date passes: the memory keeps working, and the hub’s health check lists it as overdue until you close it. Only a temporary memory can be closed this way; your everyday memories are never taken away by it.
All of it happens on the dashboard’s Memories page:
- Create one with New temporary memory next to the memory picker: a name (it becomes part of the tool names your AI tools see), what it is for, and in how many days it is due.
- Pick it in the memory picker. A banner above its notes says when it is due, or that it is overdue.
- Keep notes… lists its notes: tick the ones worth keeping, choose the everyday memory to copy them into, and copy them.
- Close… asks you to type the memory’s name, then closes it.
Starting from something, instead of nothing
Section titled “Starting from something, instead of nothing”If you’re moving from an existing setup rather than starting empty, an
importer can read your old palaia notes, or notes from other markdown-based
note tools, and bring them in with their original dates and tags intact. It
runs from the command line on the machine that hosts your hub (palaia import v2 … or palaia import basic-memory …, each with a --dry-run
that only reports); the dashboard has no import screen yet. For
developers has the technical detail on the file format itself,
for anyone who wants to script against it directly.