Definition
The working memory limit is the bounded capacity for simultaneous active maintenance of distinct information chunks in mind without external support. George Miller's classic 1956 paper proposed "the magical number seven, plus or minus two," but the modern consensus — advanced principally by Nelson Cowan since the early 2000s — places the pure capacity limit at approximately four chunks when active rehearsal strategies are blocked and chunks cannot be mutually combined or supported by long-term memory structure.
The discrepancy with Miller's estimate is largely explained by chunking: experts construct higher-order representational units that allow more raw information to be held in fewer chunks. The grandmaster who can reconstruct a chess position after a five-second glimpse is not exceeding the four-chunk limit; he is holding four chunks each comprising tactical patterns that for a novice would each occupy several chunks individually.
The limit applies to active manipulation and integration, not merely passive storage: when items must be compared, recombined, or mapped against logical or arithmetic structure, performance degrades sharply above four. The limit is largely fixed across a healthy adult lifespan, with modest effects of age and individual variation, and is closely tied to dorsolateral PFC function — the limit reflects the capacity of sustained-firing prefrontal neural assemblies. Practice produces apparent capacity gains primarily through chunking and offloading to long-term memory, not through expansion of the underlying limit.
Context
The four-chunk working memory limit is the structural reason why complex epistemic decisions are vulnerable to manipulation: any decision involving more than four simultaneously-active considerations cannot be evaluated holistically in real time and must instead be decomposed serially, leveraging long-term memory schemas, or simplified into a smaller decision proxy. Each of these strategies introduces specific vulnerabilities that kompendium techniques exploit:
- Decomposition: order-of-presentation effects allow the manipulator to control which considerations dominate the final synthesis
- Schema reliance: install or activate corrupted schemas (kompendium EA-1.4 family), and downstream decisions inherit the corruption
- Proxy substitution: substitute a tractable proxy for the actual question (kompendium EA-1.2 family), and the decision answers a question other than the one ostensibly being asked
The Decision Engine's design explicitly accounts for the four-chunk limit by structuring decision presentation so that no single human-in-the-loop step requires more than four interacting considerations. Multi-factor decisions are decomposed into auditable sequences of bounded sub-decisions, with cumulative reasoning persisted externally rather than being held in working memory. This is not a UX preference but an epistemic safety requirement.
Interactive: The Four-Chunk Ceiling
Hold a small number of items and accuracy stays near ceiling. Past Cowan's ~4-chunk limit, accuracy falls off a cliff. Chunking — combining items into higher-order patterns — effectively reduces the count, which is why experts appear to exceed the limit while staying within it. Adversarial saturation deliberately presents enough discrete items to overflow the buffer.
Discrete items presented for simultaneous maintenance
Pattern recognition that combines raw items into chunks
See Also
- Cognitive Load — Demand against capacity
- Dorsolateral PFC — Neural substrate
- System 1 / System 2 — System 2 bounded by working memory
- Mental Models — Schemas reduce effective load
- Schema Theory — Chunking mechanism
- Attentional Control — Selects what enters working memory
- Kompendium — EA technique taxonomy
Connect & Contribute
Created by Tomáš Korcak (korczis) | Open Source under GHL