Give Chat The Right Context

Use the Context Picker to set what a turn is allowed to see: which Dotabases and memories are open, which HUMAN.md applies, whether it can reach the web, and how to wall a turn down to a curated set of rows.

The Context Picker decides what an agent is permitted to look at on a given turn: which Dotabases and memories are open, whether it can reach the web, and which HUMAN.md applies.

The mechanism behind every control below is the same: selecting a source does not inject its contents into the prompt. It tells the agent where it is allowed to look. Opening a with ten thousand rows does not send ten thousand rows; it means the agent's retrieval tools may reach into it. Closing one means they genuinely cannot — a deselected Dotabase reads back as unavailable, not as discouraged. Chat covers how the turn is assembled once that aperture is set.

Open The Picker

The Context Picker is the pill in the composer's bottom row, labelled Context. Start a chat conversation covers the rest of the composer.

Its top group is four sources, each with a switch and, for three of them, a drill-down for narrowing what inside it is open.

Choose Which Dotabases Are Open

Open the Dotabases row. You get a searchable list of your Dotabases with a switch each, plus select-all and clear.

The meta line on the row reads All or N of M, and one rule governs it: a source switched on with everything inside it cleared is closed, not open. An empty selection means nothing, never everything. Selecting every item folds back to "All" rather than reading "12 of 12".

Dotabases explains what you are choosing between.

Narrow The Memory Bank

The Memory Bank row drills into three filters you can combine:

  • Maturity — fresh, maturing, or mature, so you can ask for only what has settled.
  • Memory type — the twelve types, from context snapshots and episodes through observations, patterns, decisions, and preferences.
  • Domains — the live list of domains your memories are filed under.

An "all memories" reset clears the three at once.

Let It Reach The Web

Web Access is a plain switch with nothing beneath it. It is off by default. Turning it on permits the turn to reach the web; it does not itself run a search. Leaving it off is named explicitly to the agent, so it knows the web is closed rather than merely unmentioned.

Open A Connector

Connectors is the fourth source, with a switch of its own and a drill-down whose action is Add connector; the top group offers that same action as a separate row.

Its switch rides the turn exactly as the other three do — the picker snapshots it onto the message alongside them. The Dotabases rule governs it too: a switch left on over an empty selection is closed, not open.

Choose Which HUMAN.md Applies

Below the divider, the row picks which document describes how you work for this conversation. Three kinds of choice:

  • Use global default — fall through to your account-level HUMAN.md.
  • None — an explicit opt-out for this conversation, even when a global default exists.
  • Any of your active HUMAN.md documents.

This one is per-conversation and saves immediately, which is why the row only appears once a conversation exists. If your pick points at a document that has since been archived, the row reads back as the default it fell through to.

It only applies to the direct models. Local coding agents read the real HUMAN.md off disk in your repository, so Dots injects nothing for them — a second injected copy could contradict the file the agent is already reading. In HUMAN.md Builder mode the document being authored is deliberately left out of its own prompt, so the Builder never works from a stale copy of the thing it is editing.

Set The Mode And Your Name

Two more rows sit at the bottom. Mode picks the turn's posture — Default, Ask, Plan, Research, or HUMAN.md Builder. Customize holds how you would like to be addressed and which mode new chats should start in, with a link through to the full chat settings.

Save It As Your Default

The picker's footer has Save as default and Reset. Saving writes the whole lens — sources, memory filters, web , mode — to your preferences, and every new composer starts from it. Reset is offered only when the live lens actually differs from what you saved.

Out of the box a new conversation starts with the on across all domains, maturities, and types; every Dotabase open; web access off; and Default mode.

Mention A Specific Row

Typing @ in the composer opens a picker over your rows, your connectors, and your Superpowers. Picking one inserts a chip.

A mention is a reference, not an inclusion. On send, the chip is rewritten to the row's plain title so your message reads naturally, and the reference travels separately as structured data. That reference is what draws a durable edge between the conversation and the row, and what reinforces the Memory Bank. It does not paste the row's contents into your prompt — for that, open its Dotabase in the picker so the agent can retrieve it.

Wall A Turn To A Curated Set Of Rows

Invoking a Power-Up Playbook from the Superpowers pill does something the source switches cannot: it restricts the turn's reading to exactly the rows that Playbook names, and nothing else. A Dotabase outside the Playbook's set reads back as not found even if its switch is on.

Active Playbooks appear as chips above the input. By default a Playbook applies to one message and clears after you send; pinning its chip keeps it active for the messages that follow.

The scope fails closed. A Playbook that resolves to an empty set gives the turn nothing to read, rather than falling back to your whole workspace. Invoking any other kind of — a Skill, a Prompt, a Chain — does not wall retrieval.

The Lens Is Frozen Per Message

When you press send, the whole lens is snapshotted onto that message: which sources were enabled, which were explicitly disabled, the memory filters, and the mode at send. Queued messages are snapshotted at the moment you queue them.

Two consequences follow. Changing the picker never retroactively alters a message that has already been sent or is already waiting in the queue. And every message carries a record of what the agent was allowed to see on that turn, so an answer can be checked against the sources that were actually open to it.

Give Chat The Right Context · This page in the graph

Connected Records18

Documented by
Mentions
DOT
References
18 connections.

Where To Go Next

Reading this as an agent?For Agents
Something wrong on this page?Dots On GitHub