Ask Cortexa — the vault Assistant

The Ask Cortexa tab is a conversational agent over your entire vault. Ask questions across all your thoughts, pull things out, trace how ideas connect, build reports, and set reminders — with every search and tool the assistant uses shown as a chip right in the conversation.

Enable it in Settings → LLM → Assistant. It's off by default.

Its own brain, your choice of model

The Assistant's LLM configuration is completely independent of the analysis pipeline: its own provider, model, base URL, and API key (stored in the OS keychain, never falling back to your global keys). The default is a local Ollama model — private, free, offline — and you can point it at DeepSeek, OpenAI, Anthropic, or Gemini with its own key at any time. A small local model works because the engine does the heavy lifting:

1. Vault context first — every turn runs your message through the same hybrid search ranking the app uses, so answers ground in real thoughts even if the model never calls a tool.
2. Action rounds — the model may search, read full thoughts, or call any of its allowlisted tools before answering. Tool activity is available in the collapsed activity details.
3. A streamed, guarded answer — grounded in the gathered evidence. The assistant is engineered so it can never claim it did something no tool actually did.

Model choice

For tool use, a 7B-class model (e.g. qwen2.5:7b) is noticeably more decisive than 3B models. Any cloud provider is sharper still — on its own key, without touching your pipeline config.

What it can do

• Answer across everything — "What did I decide about the token economy?" cites the actual captures.
• Reports & syntheses — "Build me a report on X" gathers widely and composes; markdown and Mermaid diagrams render in-chat.
• Time travel — period digests, "on this day", quantitative counts, opinion drift, thought lineage.
• Photo connections — ask to find images on this computer or in Google Photos. A shared connection wizard handles folder access or Google sign-in and guided photo selection, then returns private previews to this conversation. When there is one saved Google Photos account, it is preselected; multiple accounts require a choice.
• Blind spots — "What am I missing?" runs silo/open-loop/stale analysis.
• Safe actions — schedule reminders, generate diagrams, open a thought or switch views ("show me the graph"). Destructive operations are excluded by design.
• Save as thought — any answer worth keeping goes back into the vault with one click, where future searches (and future chats) will find it.

Connecting and continuing a request

Ask for what you need in the conversation. Saved connections are reused for supported read-only searches. When setup or approval is needed, Cortexa shows a short next step and a card with a named action, such as Check access. Technical arguments and results are under Action details and Recent activity.

Completing photo setup or approving a source action automatically brings an answer back into the same conversation. The follow-up uses the recorded result without rerunning the operation or making further changes. A newer question supersedes an older automatic follow-up. Previous Loom conversations remain available under Connections & history.

Google Photos still uses Google's guided selection. Photo previews and file metadata alone do not mean the assistant has analyzed the images' visual content.

Sessions

Conversations persist encrypted (the same AES-256-GCM protection as your thoughts). Each session keeps its recent history in context; the vault itself is the long-term memory, re-searched fresh every turn. Manage chats in the collapsible history rail — rename, delete (with confirmation), timestamps throughout.

For your agents: cortexa_ask

External agents can converse with the Assistant too, via the cortexa_ask MCP tool or POST /api/assistant/chat:

{ "message": "What did we decide about the relay auth model?", "session_id": "optional" }

The answer comes back with its session_id for multi-turn use — your coding agents consulting your second brain mid-task.