Spaces & tuning

One account, several libraries. Spaces keep parts of your life separate, and settings let you shape how your agent works in each one.

Spaces

A space is a fully separate library: its own pages, sources, conversations, and agent runs. Personal and work is the classic split. The agent never mixes knowledge across spaces.

Switch with the picker at the top of the library, or ⌘1–⌘9.
Each space chooses its own runtime — Claude Code, Codex or opencode — under Settings → Agent & model. That is a per-space setting; the credential it runs on is account-wide and lives under Settings → Coding agents, so you connect an agent once and then pick per space where it is used.
A space set to Claude Code gets a second dropdown there: which Claude model it runs on. The list is the models your own credential can actually reach, so it reflects your Anthropic plan rather than a fixed menu. Leave it on Default unless you have a reason — a faster model for a space you file a lot into, a stronger one for a space you ask hard questions of. Codex takes its model from your ChatGPT account and has no picker.
A space set to opencode also gets a model box, but you type into it rather than pick from a list: opencode reaches roughly 5,900 models across ~176 providers, so the value is written as provider/model — for example anthropic/claude-sonnet-4-6 or openrouter/deepseek/deepseek-v3. The provider prefix is required, and it has to be one you connected. Leave it blank to use the default.
Captures and conversations always land in the active space.

Language

Each space declares the language its knowledge is written in, under Settings → Language — so a work space can stay in English while a personal one is in Danish. It tells transcription what it is listening for and the agent what to write in, and it is worth setting: left to guess, speech-to-text mistakes Danish for Norwegian or Swedish more often than you would like.

Spaces that don't set one follow your account default (Settings → Account → Content language). Leave both unset and Memento detects the language from the audio, which is the right choice for a space you use in several languages. This is the language of your content, not of the app: Memento's own buttons and menus stay put when you switch spaces.

Instructions

Every space has a block of standing instructions you write yourself, under Settings → Instructions. House rules the agent should never need reminding of: write summaries as bullets, file anything about a particular client under its project page, always create a person page for a name you don’t recognise. They apply to everything the agent does in that space — filing a capture, running a routine, answering you in chat — and to that space alone, so a work library and a personal one can be run differently.

Two things instructions cannot change, because Memento itself reads them: the page format and the checklist grammar behind your task list. Ask for something that contradicts those and the agent follows the format and tells you why, rather than quietly breaking your tasks.

Skills

Skills are standing instructions for your agent: how you like meeting notes structured, what a “decision” page must contain, project-specific vocabulary. Enable, edit, or upload them under Settings → Skills.

Connections (MCP)

Give your agent extra tools by connecting MCP servers (calendar, task tracker, internal APIs). Define a connection once under Settings → Connections — its token is stored encrypted — then enable it per space.

Network rules

Agent sandboxes start locked down. Anything the agent tries to reach that isn’t allowed is blocked and turned into an approval request you answer under Settings → Network — allow it once and it stays allowed. A short baseline (your model provider, your library’s repository, the usual package registries) is already open so a run can work at all; everything beyond it is yours to decide.

Unlike the settings above, these rules are account-wide — one set covering every sandbox you run, not a per-space choice.

Default posture at the top of that page is the switch behind all of it, and it can be flipped. Block by default is how every account starts and is what the approval prompts belong to. Allow by default turns it around: your sandboxes can reach anything except hosts you explicitly block, and the per-host prompt goes away. That is a real loosening of the safety model — worth it if you are installing a lot of tooling and tired of approving registries, not something to leave on absent-mindedly. Internal and cloud-metadata addresses stay blocked either way, and Memento asks you to confirm before making the switch.

Account

Profile, email, and password under Settings → Account. Changing your password signs out other devices.
Your default content language sits there too, under Content language — every space that hasn't set its own follows it.
Archived chats and pages live under Settings → Archive — restore or permanently delete them there.
Space export lives with each space's settings: Markdown + originals + history, no lock-in.