A useful team needs an agreed place for facts and progress. Clero has several kinds of context, and they serve different purposes. Do not assume that every agent sees every message or that a shared agent definition combines the histories of all its projects.
Choose where information belongs
| Information | Suitable place | Why |
|---|---|---|
| A brief, specification, or accepted deliverable | Project document | A human-readable reference the team can link and inspect |
| Status rows, mappings, checkpoints, or processed-item IDs | Clero Sheet | Structured records that multiple project agents can read and update |
| External reference material | Knowledge sources | Retrieval of relevant passages within the configured scope |
| A discussion, correction, or request | The relevant conversation | Keeps the immediate exchange and its context together |
| A past project decision | Project history, with a durable decision recorded in a document | Historical recall can help locate evidence; the accepted decision stays explicit |
| Stable agent behavior or preferences | Agent instructions, a skill, or core memory | Guidance meant to outlast one task |
| An open commitment or delegated follow-up | A tracked objective or operating item | Work remains identifiable after a single response finishes |
A deadline on an objective is not an execution timer. If something must run at a future time, create a scheduled task or recurring schedule.
Give a handoff a shared reference
Instead of “the writer knows what the researcher found,” ask for an explicit handoff:
Save the research brief in this project and link it in the handoff. Include sources, uncertainties, and the decisions the writer needs. Ask the writer to read that document before drafting. Keep the tracking sheet updated with owner, state, and output link.
This makes the transfer inspectable. Team membership alone is not proof that a particular passage reached another agent. Delegation and messages explains the task and conversation routes.
Use a sheet for operational records
A sheet is appropriate when the next run must know which item was already processed, which record maps to an external ID, or which output belongs to a task. Use a stable key and clear status values.
For example, a content workflow could track one row per approved source with a source ID, draft link, reviewer, and review state. Ask each run to read that row before creating another draft. This is a useful design pattern, not a guarantee that every workflow automatically deduplicates its work.
See Sheets for the live-tested create, save, and reopen flow.
Ask for historical evidence
The current source provides project-scoped historical recall. A useful request is specific:
Find the earlier decision about the welcome checklist's audience in this project. Show the relevant message or source and distinguish the recorded decision from your interpretation.
The recall tool searches original project messages; a separate reasoning tool supports synthesis across interactions. Results can be missing or temporarily unavailable. Treat an agent's reconstructed summary as a starting point for verification, especially when the decision affects publishing, access, or spending.
Project history and a complete active conversation window are not the same thing. Flexible chats retain separate immediate context, while persistent team conversations continue over time. Neither mode means that all history is simultaneously present in every response.
Keep durable agent memory appropriate
Current agent-memory tools support listing, adding or updating, and deactivating individual core-memory items. Use core memory for stable rules or preferences that truly belong to that agent, not as an unstructured database of project status.
You can ask:
Show the durable operating rules you have saved for this agent. Identify anything specific to an old project before changing it.
If a rule is wrong, specify the exact correction and ask for confirmation of the stored change. Deactivating a memory item is not a promise to erase the original chat, documents, or service records. These memory controls are source-backed and were not live-tested in the demo.
Respect scope
An agent definition may be reusable across spaces. Its audience, permissions, and current runtime resource scope remain distinct. Sharing an agent does not automatically grant access to another space's documents, messages, or team data.
For current retrieval choices, see Agent knowledge. For access boundaries, see Data and permissions.