Plans: All plans · In the app: Settings → Context
Tabs
Overview
The Overview shows four tiles: Context health, Brand Records (trusted and pending), Evidence (indexed documents and live sources) and Active gaps. Below them:- Needs attention: up to four next steps, such as records drafted by AI that need review, missing categories, questions worth answering, records due for re-verification, failed documents, conflicts, or connecting a live knowledge source.
- Trust and freshness: structural coverage, human trust, freshness, evidence readiness and resolved decisions.
- Recent evidence: the last five documents indexed.
Brand records
Brand records are the facts every agent uses. Each record has a type, a title (up to 120 characters), a description (up to 4,000), an optional expiry date, and up to 12 detail fields.Create records
- Generate from website
- Import
- Manually
- From evidence
Reads your brand’s homepage and drafts records from it. Sites that render their content only in the browser may return little.
Trust and review
Only approved and locked records that have not expired reach Quotable Agent and routines. Everything else waits.- Records you write by hand are saved as approved. Records drafted by AI start as pending.
- Approving, locking, rejecting, deleting or editing an AI-drafted record asks for a Reason for your decision. Editing one also asks which indexed document supports the change and whether that source is primary or secondary.
- The Review queue lets you approve, lock or reject in batches, with Compare to see the change against current memory.
- Each record has a Version history with Restore. Deleted records can be restored from Recently deleted records. Roll back to a revision undoes every later change as new revisions; nothing is erased.
Review windows
Records go stale. Each type is due for re-verification after a set time since it was last verified:
Records past their window show as due for review. Approved Strategy records also expire automatically when their window passes; locked records never do.
Never-claim rules
A never-claim rule is a phrase or pattern your brand must never be credited with, such as “SOC 2 certified” before you are. Add rules in the Never claim panel of Brand Records.
While a rule is on:
- AI-drafted records that match it are dropped before they reach review.
- Approved records that match it are left out of what Quotable Agent, routines and analyses receive, and out of your llms.txt.
Never-claim rules filter the facts Quotable Agent and routines are given. They do not scan the text
an agent produces.
Other governance tools
- Supersession: replace one trusted record with another of the same type, with a reason. The old record keeps its history.
- Factual corrections: propose to update a record, mark it stale or reject it, backed by an indexed source. The record changes only after review.
- Learning loop: how often suggestions are accepted, and how answers performed by context pack. Feedback on Quotable Agent answers feeds it but never edits records directly.
llms.txt
The llms.txt button copies or downloads a brand-facts file built only from approved or locked, unexpired records that no never-claim rule blocks. Publish it at the root of your site for AI crawlers.Knowledge Base
The Knowledge Base is the evidence behind your records: call transcripts, documents, CRM notes and support tickets.The Knowledge Base, adding evidence and synthesizing records from evidence
are on every plan.
Add evidence
- Connect a source. Connectors for call recorders, document stores, CRMs and support desks are coming soon. See Knowledge sources.
- Paste text. Add evidence takes a title (up to 200 characters), an optional source URL, and between 40 and 200,000 characters of text: notes, research, a transcript.
Indexing
Every piece of evidence is indexed before agents can use it. Its status shows where it is:Authority rules
Not every source deserves the same trust. Authority rules set how much a source, a kind of document, or a specific owner can establish on its own:
Rules cap the model’s judgment by default. Tick Override the model’s answer instead of capping it to make a rule decide outright. The most specific rule wins: owner, then source, then document kind. Rules apply on the next synthesis.
Retention
- Old versions of a document lose their text 90 days after being replaced; a placeholder remains so citations still resolve.
- Evidence older than 365 days is marked stale.
- Disconnecting a source removes its indexed copies and sends AI-drafted records based on them back to review.
- Deleting a document detaches it from records, and AI-drafted records that relied on it return to pending.
Personal and team context
Brand records are facts. Personal and team context are preferences about how answers should look: which metrics you care about, how reports are formatted, what to prioritize.
Precedence runs from Brand Records, to Team Context, to Personal Context. The lower layers refine how answers are presented; they never override a fact.
Preferences have a category (Reporting, Priorities, Metrics, Communication, Workflow, Constraints, Custom), a title and a body. You can promote a personal preference to Team Context, propose either kind as a brand record, export to CSV, or delete them. The Precedence and conflicts panel shows what is in your current Quotable Agent context and what was left out, and why.
Intelligence
- Context gaps: questions synthesis could not answer from evidence, ranked by how much the answer would matter. Answer (and store the answer as a brand record, team context or personal context), Skip, Snooze for a day, a week or 30 days, or Assign someone and send them the question.
- Atomic fact graph: each fact as a statement with subject, predicate and object, its authority, confidence, validity dates and evidence.
- Conflicting evidence and Memory relations (supports, contradicts, supersedes) that you confirm or reject.
- Synthesis activity: recent runs, with Run again, Cancel and Stop after this batch.
- Fact history with restore.
- Perception gap and Prompt coverage analyses, run on demand.
Roles
Viewers can read the Context Manager. Everyone else with access to the brand can edit records, evidence, rules and gaps. See Team and roles.Next steps
Knowledge sources
Connect call recorders, documents, CRMs and support desks.
Skills
Write reusable instructions for Quotable Agent and routines.
Quotable Agent
Ask questions grounded in this context.