Context, memory & approval

How continuity works

An agent receives instructions, selects relevant files, and uses their contents for the current task. You review important results and save the accepted facts. A later session must load those saved records again.

This is file-based context retrieval. It does not change the model’s training, guarantee automatic loading, or make every stored statement true.

Give each record one job

Record	Keep here	Avoid
Project context	Objective, scope and entry points	A complete chat transcript
memory.md	Approved decisions, durable constraints, terminology	Temporary status and unapproved proposals
handoff.md	Current state, validation, blockers, exact next action	Years of accumulated session logs
DECISIONS.md	Decision record and approval status where used	Silently changing proposed into accepted
Task/action note	Owners, due dates, work items	Facts only present in chat

Review before approval

Ask the agent to show the exact proposed changes. Check evidence, scope and unresolved questions. Authorize only what you accept. Save open questions as open. A general request to summarize is not approval for a new architecture decision.

Community templates can need explicit access and review metadata. Pro has a paired context lifecycle: complete and review the generated files before using the provided activation command. Do not mix those procedures.

Keep context bounded

Load the root router and the current project’s relevant records. Bring in a source document when the task needs it. Avoid broad ingestion of archives, reports and attachments without a reason and permission.

Check whether it worked

Inspect the saved files, then start a new session. Ask the agent to name the files it used and recover a known decision and next step. Correct a fact through an approved edit and repeat. The quickstart provides a concrete acceptance exercise.