Working Memory

Working memory has a severely constrained capacity, famously quantified by Miller's Law as 7 ± 2 items. However, the effective unit is the 'chunk'—a meaningful grouping of information. For example, the letters F-B-I-C-I-A are seven items, but when chunked as 'FBI' and 'CIA', they become two manageable units. In AI systems, this is mirrored by context window limits in language models, where strategic structuring of prompts (chunking) maximizes the utility of available tokens.

Information in working memory decays rapidly unless actively maintained through rehearsal loops. The phonological loop rehearses auditory information (like repeating a phone number), while the visuospatial sketchpad maintains visual images. In artificial agents, this corresponds to state tracking and hidden state vectors in recurrent neural networks (RNNs) or transformers, where attention mechanisms and recurrent connections actively maintain task-relevant information across processing steps to prevent it from fading.

The central executive is the supervisory system of working memory. It does not store information but directs cognitive processes:

• Coordinates the phonological loop and visuospatial sketchpad.
• Switches attention between tasks and mental sets.
• Inhibits irrelevant stimuli or prepotent responses.
• Integrates information from long-term memory. In agentic AI, this function is performed by orchestration layers, planners, and controllers that manage sub-modules, allocate computational resources, and gate information flow based on current goals.

The episodic buffer is a temporary, limited-capacity store that integrates information from different sources into a unified, multi-dimensional representation or 'episode'. It binds data from:

• The phonological loop (language)
• The visuospatial sketchpad (imagery)
• Long-term memory (semantic knowledge, past episodes) This creates a coherent model of the current situation. In AI architectures, this is analogous to a fusion module in multimodal systems or a working memory tensor that combines embeddings from text, vision, and retrieved knowledge into a single, updated context for the next reasoning step.

Beyond passive storage, working memory's critical function is the active manipulation of information to serve ongoing goals. This includes:

• Mental arithmetic and problem-solving.
• Reordering items (e.g., alphabetizing a list in your head).
• Updating contents as new information arrives.
• Synthesizing new ideas from held elements. This manipulation is the core of reasoning. In AI, this is implemented through chain-of-thought prompting, where a model holds intermediate reasoning steps in its context, and algorithmic loops within agents that transform and update their internal state representations to subserve task execution.

Working memory is highly vulnerable to interference, where similar information disrupts the retention of target content. Key types include:

• Proactive Interference: Old memories interfere with learning new ones (e.g., remembering an old phone number instead of a new one).
• Retroactive Interference: New learning disrupts recall of older information. This fragility necessitates robust distractor suppression mechanisms. In machine learning, this is seen in the catastrophic forgetting of sequential learning and the challenge of maintaining task-specific context in a transformer's attention mechanism when processing long, noisy sequences. Techniques like attention masking and rehearsal buffers are engineered countermeasures.

The central executive is the supervisory component in Baddeley's multi-component model of working memory. It is responsible for:

• Controlling attention and focusing on relevant information.
• Coordinating the subordinate 'slave systems' (phonological loop, visuospatial sketchpad).
• Switching between tasks and retrieving information from long-term memory. In AI architectures, this function is often simulated by a controller or planner module that allocates computational resources and manages sub-processes.

The episodic buffer is a later addition to Baddeley's working memory model. It acts as a temporary, limited-capacity storage system that:

• Integrates information from the phonological loop, visuospatial sketchpad, and long-term memory.
• Binds these disparate elements into a unified, multi-modal episode or conscious experience.
• Serves as an interface between working memory and long-term memory. In agent design, this is analogous to a context window or short-term memory module that creates a coherent narrative from recent perceptions, actions, and retrieved facts.

Cognitive load refers to the total amount of mental effort being used in the working memory system. It is influenced by:

• Intrinsic Load: The inherent complexity of the information or task.
• Extraneous Load: How the information is presented (poor design increases load).
• Germane Load: The effort devoted to processing, constructing, and automating schemas (learning). For AI agents, this translates to managing the context window budget, where excessive tokens (details, instructions, history) can overwhelm the model's capacity, leading to degraded performance on the core task.

Dual-task interference is the performance decrement observed when two tasks are performed concurrently, because they compete for a limited pool of shared cognitive resources. Key characteristics:

• Occurs when tasks require the same modality (e.g., two verbal tasks) or the same central executive processes.
• The degree of interference is a key measure of working memory capacity.
• In AI systems, this manifests when an agent attempts to multiplex between unrelated sub-tasks without sufficient parallelization or rapid context-switching mechanisms, causing errors or slowdowns.

This dichotomy describes two modes of operation for cognitive systems:

• Controlled Processing: Slow, effortful, capacity-limited, and requires conscious executive attention (working memory). Used for novel, complex, or non-routine tasks.
• Automatic Processing: Fast, effortless, parallel, and operates outside of conscious control. Developed through extensive practice (e.g., reading, driving). In agentic AI, controlled processing is analogous to the main reasoning loop, while automatic processing can be seen in cached responses, fine-tuned sub-models, or tool use that has become highly practiced and efficient.

The Supervisory Attentional System (SAS) is a central component of the Norman and Shallice model of executive control. Its functions include:

• Intervening in non-routine situations where automatic, schema-driven responses (controlled by 'contention scheduling') are insufficient or inappropriate.
• Planning and decision-making when novel sequences of action are required.
• Troubleshooting and overcoming habitual responses. This is a key architectural blueprint for AI agents, directly inspiring the need for a high-level planner or orchestrator that overrides simple stimulus-response patterns to achieve complex, novel goals.