Is internal state the same as a database?

No. A database may store it, but the concept describes the information used by the agent rather than a particular storage technology.

Must it contain probabilities?

No. Some agents use deterministic variables or symbolic facts; probabilities are useful when uncertainty is modeled explicitly.

Can the same observation produce different actions?

Yes. Different retained information or goals can make different responses appropriate even when the latest observation is identical.

AI Glossary / Graphs and symbolic AI / Beginner

Internal state

Definition

An internal state is information an agent maintains within itself to support decisions over time. It can summarize earlier observations, previous actions and estimates of the current situation, especially when the latest observation does not reveal everything relevant.

What it means

An agent's internal state is different from the actual state of its environment. The environment might contain an open door, while the agent's stored estimate still says the door is closed. Internal information can be incomplete or wrong.

It also differs from a single current observation. A sensor reading tells the agent something now; internal state can combine that reading with what happened earlier. Two identical readings can therefore lead to different actions when the retained histories differ.

The term is broad. In agent design it often refers to a maintained representation of the situation. In other computing contexts it can describe program variables or hidden numerical representations. This entry focuses on the decision-making agent meaning.

How it works

A model-based agent starts with an initial internal state. After an action and a new observation, an update procedure revises that state using assumptions about how the world changes and how observations relate to it.

The decision procedure then uses the revised information to select an action. This update-and-decide cycle continues as the agent interacts with its environment. The retained information should be relevant to future choices rather than merely a large archive of everything encountered.

Under uncertainty, the representation may contain alternatives or probabilities instead of one asserted situation. In a correctly specified POMDP model, a belief distribution over possible environment states summarizes the history for subsequent decision-making. A casually chosen summary does not automatically have that property; discarding relevant history can make decisions less reliable.

Three practical examples

A mobile agent remembers which corridor branches it has already explored when the current camera image cannot reveal that history.
A workflow assistant records that a request is awaiting confirmation rather than treating its last attempted action as completed.
A simulated monitoring agent combines an earlier estimate with an imperfect sensor reading to revise its belief about a hidden condition.

Think of it this way

Think of a navigator's working notebook. The view through the window is the current observation; the notebook holds the route, previous turns and unresolved uncertainty about the location.

The notebook can help interpret the view, but an incorrect entry can also mislead. The analogy concerns retained information, not a claim that an artificial agent experiences memory or awareness like a person.

Worked example

Updating a belief instead of treating a reading as certainty

Suppose an agent monitors whether a fictional cabinet is open or closed. Before a reading, its internal belief assigns probability 0.4 to open and 0.6 to closed. Assume the cabinet does not change during this example.

A sensor reports open with probability 0.9 when the cabinet is open and probability 0.2 when it is closed. The joint weight for open and an open report is 0.4 × 0.9 = 0.36. The corresponding closed weight is 0.6 × 0.2 = 0.12.

After receiving an open report, the total weight is 0.48. The updated probability of open is 0.36/0.48 = 0.75, leaving probability 0.25 for closed. The report supports open, but does not establish certainty.

If a second open report is conditionally independent of the first given the unchanged cabinet state, the new weights are 0.75 × 0.9 = 0.675 and 0.25 × 0.2 = 0.05. Normalizing gives 0.675/0.725, approximately 0.931, for open.

The independence and unchanged-state assumptions matter. Repeatedly processing the same reading as new independent evidence would create unjustified confidence. If an action could move the door, the update would also need a transition model before incorporating the next observation.

Illustrative case study

The fictional Alder Agents team tests an assistant that requests a change and immediately records the change as successful. When a tool call times out, the assistant's internal record can disagree with the external system.

Reviewers introduce separate pending, confirmed and uncertain statuses, together with the evidence supporting each update. They test delayed responses and repeated observations so that one result is not accidentally counted twice. The next action depends on the recorded uncertainty as well as the latest message.

The team also checks when old information should expire. This fictional case illustrates the importance of state semantics and update rules; it does not prescribe a particular software architecture or report measured improvements.

Common misconceptions and mistakes

Internal state is a perfect copy of the world. It is a representation that can be uncertain or stale.
More stored history always means better decisions. Irrelevant or incorrectly interpreted information can add cost and confusion.
A successful action request proves the intended external change occurred. Confirmation and model assumptions still matter.
Any compact memory is a sufficient summary. The summary must retain the distinctions needed for the decision problem.

Business use

The concept is useful when reviewing assistants, robots and automated workflows that operate across multiple steps. Ask what the system believes, why it believes it and what event changes that belief.

Practical reviews should distinguish observations, predictions, attempted actions and confirmed outcomes. Test interruptions and conflicting updates, because these often expose hidden assumptions about continuity. A retained status should have a clear meaning that later components can use consistently. For uncertain situations, making uncertainty explicit can support a better next observation or a request for human input instead of an unsupported conclusion.

Check your understanding

Why should the agent not process one sensor report twice as two independent observations?

Answer: The second processing adds no new evidence. Treating it as independent can increase confidence without justification and distort the internal belief.

Related terms

Agent
Percept
Belief state
Transition model
model-based reflex agent
state estimation
partial observability
observation model

Learn next

Choose a three-step workflow and list what an agent must remember after each step. Separate confirmed facts from predictions, then describe how a timeout or contradictory observation should change the internal record.

Agent

Sources

AIMA's author algorithm collection gives the model-based agent's state-update cycle: https://aima.eecs.berkeley.edu/algorithms.pdf

The author slides distinguish agent architectures and retained state: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter02.pdf

Cassandra explains belief distributions for partially observable models: https://www.pomdp.org/tutorial/pomdp-background.html

The cabinet probabilities and Alder Agents case are original illustrations with explicit assumptions, not observations from a deployed system.

Previous: Internal Schema Next: Internal validity