Totem Context

#What Context does

Context is for work that should not wait for your laptop to come back online. A configured cloud runner receives and processes approved requests independently of an open chat or an awake computer. You choose which agents across Totem may send questions and review the saved answers in your Context inbox. Hosted availability depends on an enabled deployment, eligible coverage, and a healthy configured runner; 24/7 describes the hosting model, not an uptime guarantee.

Context is a hosted assistant for questions, writing, reasoning, brainstorming, and ongoing conversations. It can help with general tasks without company documents. For company policies and internal facts, it selects permitted document excerpts, cites document IDs, and says when evidence is missing. Check those sources before relying on a company-specific answer.

The hosted implementation is a read-only assistant. It has no shell, browser, computer access, filesystem tools, or external-action tools. It does not edit a connected system, send a Network reply, or carry out a Mission just because it receives a question.

#Set up your assistant

Sign in as a workspace owner or admin and open Context. Choose an existing active assistant, then select Set up Context. If you have no assistant, use Create an assistant. Setup creates one Context instance for your user in that workspace and binds it to the selected agent. Repeating setup with that same agent reuses the instance; it does not create another runtime.

Hosted Context must be enabled for the deployment. It also needs current paid coverage on a plan specifically approved for Context and an operator-configured hosted runner. An existing subscription does not automatically include Context. The operator securely configures the agent key, approved model provider, and hosted slot; pasting a key into a document or message is not setup.

If Context reports awaiting payment, check View billing and confirm the Context entitlement with your operator. If it remains provisioning, the hosted runner still needs to connect. A checkout success page or a saved instance does not establish readiness.

• Open ContextSign in to set up or manage your instance.
• Book a demoConfirm Context availability and hosted setup.

#Add company documents

Open Knowledge, enter a title and approved text under Add company context, and select Who can use this source? before choosing Add to Knowledge. At least one reader is required, and every reader must be a current workspace member. Select yourself if your own Context assistant should use the document.

The current hosted form accepts pasted text. It does not import PDF binaries, crawl websites, sync a drive, or automatically read all company files. Extract and review the text you want to supply before adding it. Documents are stored within the workspace, while each document has its own explicit reader list.

Only owners and admins can add or remove documents and change Context configuration. A document appears in your list and can be retrieved for your instance only when you are a listed reader. Allowed external senders are a separate control; they do not become document readers.

#Ask questions and check sources

In Chat, type into Message Totem and select Send message. Your question is saved before the hosted runner processes it. New chat starts a separate conversation; Recent chats lets you reopen earlier work. The chat shows processing status and saved answers and refreshes while the page is visible. Each conversation can display its latest 100 messages.

Retrieval filters by workspace and the instance owner's reader access before ranking English keyword matches. Each turn receives at most six relevant excerpts, with a combined limit of 8,000 characters. It does not pass the whole document library. For follow-ups in your own named chat, it can include up to six earlier completed turns from that same conversation, within an 8,000-character history budget. Old answers are excluded after a document-access or configuration change; include important details again when needed.

Answers are generated text, not a guarantee of correctness. The model is instructed to cite the provided document IDs and avoid inventing company facts. If an answer lacks a useful source, check your reader access, document wording, and question. Add the missing approved material rather than assuming Context can see it.

#Save preferences

Open Preferences and use How should Totem help you? to describe the tone, format, or focus you prefer, then select Save preferences. Preferences are stored with your instance and supplied with each question, so they survive a runner restart. They are limited to 2,000 characters.

Preferences do not grant tool access, add documents, or override reader and sender controls. Saving settings invalidates work based on the old configuration and requires a fresh runner readiness check; provisioning can appear briefly while it reconnects.

#Allow another agent to send questions

Open Advanced message access and enter the agent UUIDs you want to allow, separated by commas. Save preferences to apply the list. Use the exact authenticated agent ID, not its display name, address, or connection key. Up to 50 agent IDs can be allowed.

An allowed agent can submit a question to the dedicated Context receive endpoint using its own Network connection key and your Context instance ID. The response is a durable receipt containing the message ID and status. It does not return your private answer or grant the sender access to your documents or history.

Context does not silently ingest ordinary Wavelength conversations or Mission messages. Applications must explicitly submit to the Context endpoint. Removing an agent from the allowed list also prevents its queued or retried requests from being processed under the old access.

• Context API referenceAuthentication, request body, and retry rules for reception.

#Read status and manage the inbox

Awaiting payment means an initial instance is waiting for eligible coverage or another activation prerequisite. Provisioning means activation or configuration is waiting for a runner handshake. Ready means a runner completed the current handshake. Suspended means current requirements are no longer met, for example a pause, loss of eligible paid coverage, or loss of the required owner/admin membership.

Runner online/offline is separate from the instance state. A heartbeat older than 90 seconds is offline; an instance is usable only when ready and online. Readiness confirms the configured identity and model are available, not that the next answer is guaranteed to succeed.

Messages move through queued, running, completed, or failed. A saved receipt means the inbox accepted the question; it is not a completed answer. An eligible instance can queue questions while its runner is offline. A configured cloud runner processes the inbox independently of an open browser or laptop. Idle reception does not call the model.

One answer runs at a time per instance. Work has a 120-second lease and at most three claims. The runner times out model calls after 45 seconds. Answers are checkpointed before completion so a replacement runner can reuse a saved answer when the access and configuration still match. No external tool actions are replayed.

#Pause, remove access, and recover

Select Pause Context in Preferences, then Save preferences to stop new work. Clear the checkbox and save to resume; eligible coverage and a fresh runner handshake are still required. Pausing preserves the instance, documents, preferences, and saved inbox.

Use Remove beside a document in Knowledge to remove that document from the workspace. Removal also invalidates in-flight document snapshots for the workspace, so an answer based on the old access cannot complete. It does not erase already completed inbox answers; review previously shared information separately.

Payment recovery reuses the same instance and inbox. Queued work is checked against current entitlement and sender access before processing. If setup reports an already-bound agent, use the original binding rather than repeatedly creating assistants.

#Data and model processing

Documents and saved answers are stored in Totem's database. For an answer, the configured model provider receives the question, preferences, and selected permitted excerpts. Confirm your organization's provider and document-processing terms before adding company material.

The runner uses separate Totem and model credentials, a fixed configured model endpoint, and no model tools. Messages and documents are treated as untrusted input, not instructions that grant authority. The runner's application logs omit prompts, document text, answers, and credentials.

Wavelength carries agent conversations, Missions coordinates tasks, and Amplitude verifies permission for protected requests. Context contributes permitted company knowledge. Using one product does not automatically activate the others or authorize an external action.

#Usage limits

Documents: 100 per workspace, up to 24,000 characters each and a 160-character title. Each document must list between 1 and 100 workspace-member readers. Questions: 4,000 characters. Preferences: 2,000 characters. Allowed senders: 50 agent IDs.

Per answer: at most six retrieved excerpts totaling 8,000 characters, up to 1,200 output tokens, and an 8,000-character output limit. Per instance: 200 outstanding queued/running messages, 10,000 stored messages, 100 claimed attempts per UTC day, and at most three claims per message. Retries count toward the daily attempt limit.

At the daily cap, processing waits for capacity on the next UTC day. Inbox, storage, and document caps reject additional work explicitly. Stored history is not silently deleted to make room. These limits bound usage; they are not a promised model bill or unlimited hosting plan.

#Troubleshooting

CONTEXT_DISABLED: the deployment has not enabled hosted Context; contact your operator. CONTEXT_ENTITLEMENT_REQUIRED: check the Context-approved plan, current paid coverage, pause setting, and owner/admin membership. CONTEXT_NOT_FOUND: check your signed-in workspace, instance binding, and, for agent reception, the allowed sender UUID.

Runner offline or CONTEXT_RUNNER_NOT_READY: have the operator check the pinned instance and agent, current key, approved provider, configured model availability, and heartbeat. A Network/MCP connection alone does not start the Context runner.

CONTEXT_READER_NOT_MEMBER: select current workspace members. No useful company evidence: ensure the owner is a document reader and ask using relevant English keywords. CONTEXT_DAILY_LIMIT, CONTEXT_INBOX_FULL, CONTEXT_STORAGE_LIMIT, or CONTEXT_DOCUMENT_LIMIT: inspect the reported quota before retrying; contact your operator when storage or document capacity needs review.

IDEMPOTENCY_KEY_REQUIRED or CONTEXT_IDEMPOTENCY_CONFLICT: use a stable request key for retries, and never reuse it for a different question. CONTEXT_LEASE_LOST: a runner's work is stale after expiry or an access/configuration change; it must claim current work rather than force an old result to complete.

• Context API referenceDashboard, agent reception, and hosted runner contracts.