Context and memory

An agent has a continuing Codex conversation, persistent files, and access to shared knowledge. These serve different purposes.

Conversation context

The runtime resumes the agent's saved conversation across work. Channels notifications, Cortex work, and scheduled tasks can enter that shared context. They are not separate, isolated model sessions for each issue or conversation.

A new event can steer an active turn. The agent must distinguish which conversation or task each instruction belongs to and communicate results in the right place.

Durable memory

Files in /home/agent persist beyond a turn or computer restart. The home template tells the agent to record useful decisions, unfinished work, and reference notes in its memory and documentation folders.

Persistence does not mean every file is included in every model request. The agent must read the relevant material. Long conversations may also be compacted; do not rely on verbatim recall of everything previously said.

Shared knowledge in Cortex

Use Cortex for knowledge other people and agents should retrieve and maintain together. Use local files for working state and notes specific to that agent.

For an ongoing responsibility, link the governing modules and ask the agent to consult them before acting. Its identity must have access to those modules.

Keep boundaries explicit

Shared context is not a separate security boundary for every conversation. Give the agent only the access appropriate for its role, and specify where results may be shared.

See Identity and access.