Follow material in an order that builds understanding.
inAiLoading page content...
Unit 04 · information over time
Plan for this page
After this unit, you should be able to
Terms you will use
These definitions prepare you for the reading; you do not need to memorize them.Information available to a model or controller for the current operation, whether supplied by a person, retrieved, or produced earlier.
The structured record of progress, decisions, results, pending steps, and constraints needed to continue one piece of work.
Information stored beyond the current interaction or task so it can affect later behavior, subject to access, retention, correction, and deletion rules.
A model can use text in the current request, recent messages, retrieved documents, and tool results without storing them for the future. This is temporary context: available now, but not necessarily retained after the session or operation ends.
People may experience conversational continuity and assume the system remembers everything. The interface must not substitute for evidence. Ask what is actually sent into each operation, where it came from, how long it remains, and whether later tasks can retrieve it.
A multi-step task needs facts such as completed actions, current status, approvals, remaining budget, failures, and the next milestone. Structured task state is often more reliable and inspectable than a long transcript because each field has a purpose and an update rule.
State can become inconsistent. A booking may be cancelled in one system while an old task record still says confirmed. Good design records source, timestamp, version, and confidence where needed, and refreshes volatile facts rather than treating every stored value as timeless.
Cross-session preferences or history can reduce repeated work, but persistence changes the privacy and control question. Who may write or read the state? Can the person inspect, correct, export, and delete it? Is the stored information necessary for the stated purpose?
More memory is not automatically better. Old, inferred, sensitive, or incorrectly attributed information can shape later actions in ways the person cannot see. Minimize retention, separate task records from long-term preferences, and define expiration and review for information that changes.
Interactive model
Inputs and retrieved material available for one model or controller call.
Misconception to correct
Active checks
Explain back and transfer
Source context
Provides a practical route into observations and iterative task state; its implementation vocabulary should not be treated as a universal architecture.
Supports analyzing data processing, purposes, privacy risk, governance, and control when context becomes durable; it is not specific to agent memory.
Supports mapping data, context, actors, monitoring, and lifecycle risk across an AI system rather than assuming that more retained information is always beneficial.
