Skip to content

Agents: memory & learning

Agents get better over time through a durable persona, automatic learning, and retrospectives from completed work.

Capabilities

Role prompt

A structured, durable persona (Role, Rules, Examples, Sources) loaded into every run and editable via agent settings or tools.

Automatic role learning

A background agent proposes role-prompt additions from teaching moments in chat, which humans accept or reject. Enabled by default; controlled by a feature flag.

Task memory retrospectives

Closed tasks generate memory snippets that are injected as prior experience into similar future runs.

Agent personalization settings

Per-agent model, reasoning effort, service tier, Codex model/tier, default skillsets, and spend limit.

Last updated 2026-07-04