SOAR Working Memory: Active Context in the Decision Cycle

SOAR working memory is the system's active context during a decision cycle.

It holds the current state, the selected operator, relevant objects, and goal context: everything the architecture needs to decide what to do next. It is not long-term storage. Elements remain while supported and leave when they are retracted, explicitly removed, or no longer linked to a state.

Working memory elements (WMEs)

SOAR represents everything in working memory as working memory elements, usually called WMEs.

A WME is a three-part structure: an object identifier, an attribute, and a value.

WME field	Example and meaning
Identifier	`S1`: the state or object being described
Attribute	`task`: the property being recorded
Value	`password-reset`: the current value of that property

Several WMEs combine to describe the same object. A task in working memory might look like:

text

S1 ^goal        resolve-support-issue
S1 ^user        Susan
S1 ^task        password-reset
S1 ^last-result failed
S1 ^operator    ask-for-failure-code

This simplified sketch is not complete Soar syntax. It shows a structured active state that production rules can match against.

Working memory is active context, not storage

An AI agent may have access to policies, notes, past conversations, tool logs, and user profiles.

But it cannot act on all of them at once. It needs a current state, a surface that makes the present problem inspectable and actionable.

SOAR working memory is that surface. Production rules match against it, and operator-application rules update it. Retrieval from semantic or episodic memory adds new elements when the system needs long-term knowledge.

The design principle is the same one ACT-R enforces with buffers: knowledge in storage is not the same as knowledge in play. Only what enters working memory is active.

SOAR working memory vs other memory types

Memory	Role and persistence
Working memory	Active context; WMEs remain while supported or until explicitly removed
Procedural memory	Long-term production rules that match against working memory
Semantic memory	Long-term general knowledge retrieved into working memory
Episodic memory	Long-term episodes captured from working memory and later retrieved with cues

Why this matters for AI agents

Without a clean current state, every action requires re-reading raw input, guessing what is still relevant, and re-deriving the task context from scratch.

With a working memory, the system can ask precise questions at each step:

What is the goal?
What changed since the last operator applied?
What operator is currently selected?
What long-term knowledge is now active here?

Those questions make agent behavior easier to inspect, debug, and extend.

The takeaway

SOAR working memory is not a cache or a context window. It is a structured, dynamic representation of the problem being solved across decision cycles.

For AI builders, the lesson is that an agent needs a clean current state before it can choose a reliable next action.

Sources

Go back to What Is SOAR Cognitive Architecture?, continue with Operators in SOAR, or return to the Concepts hub.

What Is SOAR Working Memory?