Working Memory
A concept in cognitive psychology referring to the human capacity to simultaneously process and store information. Its limited capacity, approximately 7±2 information chunks, is closely related to the cognitive load imposed on developers by code complexity.
Core Facts
Psychologist George Miller proposed the "magical number 7±2" in 1956, revealing the bottleneck of human working memory.
Classic research by psychologist George Miller shows that human working memory can only process 7±2 chunks of information at a time.
Human working memory capacity is limited, typically around 4±1 chunks.
Timeline (last 90 days)
Working memory is a temporary storage system with limited capacity, responsible for maintaining and manipulating information within seconds, with the prefrontal cortex as its core brain region.
Birdwatching requires mobilizing cognitive functions such as attention, working memory, pattern recognition and spatial orientation simultaneously.
Long loops easily blow out the context window and also drive up API costs. A common approach is to maintain a lean "working memory" that keeps only file snippets relevant to the current task and the most recent execution results. In addition, a maximum iteration count is set to prevent the agent from falling into an infinite loop.
Context-Mode output sandboxing refers to policy-based interception and filtering of tool outputs, completing the processing before writing into the agent's working memory to prevent unnecessary tokens from entering the context.
The key to output sandboxing is to place the processing action before writing into working memory, equivalent to setting a gate at the entrance of the context window rather than cleaning up afterwards.
Mistakenly storing working memory in a long-term vector store will degrade recall quality.
LangGraph's graph state mechanism is naturally suited to carrying working memory, and the state passed between nodes is the concrete form of working memory.
Memory management means deciding what the agent carries forward, how long it carries it, and what it discards; working memory holds the current task and is discarded when the task ends, while persistent memory holds durable facts and is selectively retrieved.
Production-grade agent memory architecture can be divided into three layers: working memory (active context), persistent state (checkpoints and recovery), and long-term memory (cross-session retrieval), each with different storage, retention, and access patterns.
Human working memory capacity is limited, typically around 4±1 chunks.
15 more timeline events
All Facts (20)
Psychologist George Miller proposed the "magical number 7±2" in 1956, revealing the bottleneck of human working memory.
80%VerifiedClassic research by psychologist George Miller shows that human working memory can only process 7±2 chunks of information at a time.
80%VerifiedHuman working memory capacity is limited, typically around 4±1 chunks.
65%VerifiedWorkBuddy's generated conversation memory feature automatically organizes and records context, enabling the AI to gradually understand the user's preferences and habits.
65%VerifiedCognitive science research indicates that human working memory can only process 4±1 chunks of information at a time
65%VerifiedCowan's embedded-processes model holds that the effective capacity of working memory is closer to 4 chunks.
65%VerifiedPersistent memory is typically divided into three levels: short-term working memory, episodic memory, and semantic memory
65%UnverifiedNelson Cowan revised the working memory capacity figure to a more conservative 3 to 5 chunks of information in 2001
60%UnverifiedWorking memory is a temporary storage system with limited capacity, responsible for maintaining and manipulating information within seconds, with the prefrontal cortex as its core brain region.
50%UnverifiedBirdwatching requires mobilizing cognitive functions such as attention, working memory, pattern recognition and spatial orientation simultaneously.
50%UnverifiedLong loops easily blow out the context window and also drive up API costs. A common approach is to maintain a lean "working memory" that keeps only file snippets relevant to the current task and the most recent execution results. In addition, a maximum iteration count is set to prevent the agent from falling into an infinite loop.
50%UnverifiedThe key to output sandboxing is to place the processing action before writing into working memory, equivalent to setting a gate at the entrance of the context window rather than cleaning up afterwards.
50%UnverifiedContext-Mode output sandboxing refers to policy-based interception and filtering of tool outputs, completing the processing before writing into the agent's working memory to prevent unnecessary tokens from entering the context.
50%UnverifiedMistakenly storing working memory in a long-term vector store will degrade recall quality.
50%UnverifiedLangGraph's graph state mechanism is naturally suited to carrying working memory, and the state passed between nodes is the concrete form of working memory.
50%UnverifiedMemory management means deciding what the agent carries forward, how long it carries it, and what it discards; working memory holds the current task and is discarded when the task ends, while persistent memory holds durable facts and is selectively retrieved.
50%UnverifiedProduction-grade agent memory architecture can be divided into three layers: working memory (active context), persistent state (checkpoints and recovery), and long-term memory (cross-session retrieval), each with different storage, retention, and access patterns.
50%UnverifiedA managed agent runtime typically includes capabilities such as task queuing and scheduling, cross-step state persistence, error retry and checkpoint-based resumption, resource management, and observability.
50%UnverifiedWorking memory is the scratchpad for active tasks, storing immediate tool calls and execution state, and is cleared when the session ends.
50%UnverifiedIn pure voice interaction, when information arrives in rapid bursts, the human brain's working memory cannot keep up, making it a suboptimal way to convey information.
50%