Context & memory

Add context intentionally#

Open Settings → Context. Personal context helps your agent understand your role, preferences, and responsibilities. Organization context describes the business and is managed by administrators.

Good context is stable and useful across tasks: what your company does, who you serve, the terminology you use, and how you prefer to work. Keep changing records such as orders and task status in their apps.

I coordinate purchasing. Present supplier comparisons in a table, keep quoted currencies unchanged, and call out missing lead times before recommending a supplier.

Personal and organization information#

Your own context is available in the personal part of the page. Organization sections may be hidden if you are not an administrator. Check which section you are editing before saving: a personal preference and a company-wide instruction have different audiences.

Role knowledge, personality, and files#

The agent's General settings separate how it behaves from what it should know about its role. Use personality for tone and behavior, and knowledge for domain facts or standing instructions. Company-wide information belongs in organization context; a single task's requirements belong in your request. The agent also fills these in as you tell it about its job in the chat, and says so in a short line; check and edit what it saved in General.

Saved workspace files and memory are also different. A file holds concrete source material or an artifact. Memory holds retrievable facts. Your chat is a third record: the agent works from a summary of earlier messages plus the latest ones, and looks up exact wording, dates, or files from further back when you mention them. Point the agent to the relevant source when exact details matter.

Workspace timezone#

Agents use the workspace timezone selected during setup or in organization settings to interpret “today,” “tomorrow,” and times without an explicit timezone. Each new request receives a fresh reference time and the saved workspace timezone, including in your ongoing chat. Specify another timezone in your request when needed.

Work already in progress keeps its original date reference, including delegated tasks and approval resumes. New recurring schedules save their resolved timezone; changing the workspace timezone does not move existing schedules.

Agent rules#

Rules are short instructions an agent follows on every message, such as payment terms for a customer or something it must never do. When you correct the agent in a chat or tell it how something must always be done, it saves a rule and says so in one short line. Find them under Rules in the agent's settings, where you can add, reword, or remove any of them. A number beside Rules shows suggestions waiting for your approval.

On a specialist agent other people use, everyone with access can shape how it works: its personality, knowledge, rules, skills and automations, from its chat or its settings. Everything else about the agent (its access, approvals, model, name and look) stays with its owner and administrators, and its members don't see those settings at all; the owner sees them marked Owner only. The owner can turn members' editing off under Access; from then on a member's correction is saved as a suggestion that applies once the owner or an administrator approves it. On someone's first agent, which acts with that person's authority, only its owner or an administrator changes anything. The agent only saves what people tell it directly, never instructions found in an email, document, or web page.

Agent memory#

Agents can save and retrieve useful facts through their memory capability. Ask the agent what it remembers about a topic, tell it when a fact is out of date, and ask it to correct or forget that fact.

Memory is useful background, but important policies and business decisions should also have an authoritative home, such as a shared document in an installed app. Point the agent to that source when asking it to apply the policy.

Keep context useful#

Review it as responsibilities change. Remove contradictory instructions, outdated process descriptions, and details the agent no longer needs. For one-off task constraints, put them in the conversation instead.

If the agent uses the wrong assumption, identify the incorrect fact and the correct source. After a major change, say in the chat what changed and which earlier details no longer apply: the chat keeps its history, so the agent needs to know what to set aside. Update context or the agent's settings wherever the old version is saved.