Hierarchical Memory
A memory architecture in agent context management that distinguishes short-term conversation from long-term knowledge, applying different eviction strategies to each so as to operate within a hard token budget.
Core Facts
Timeline (last 90 days)
The user profile uses a layered memory architecture divided into three layers: permanent preferences, situational physiological information and immediate conversation context.
The solution uses a layered memory architecture: a concise "persona" stays resident in context, while a dynamic "diary" does not enter the context and is retrieved on demand via search.
Layered memory divides context into three layers: the always-loaded core state, the on-demand-loaded architecture documents and API specifications, and the background full-text-search-style history records.
Mitigation solutions for long-term memory include hierarchical memory (hot storage and cold storage), TTL mechanisms borrowed from caching systems, and weight decay based on the forgetting curve; these solutions are still in an experimental stage and have not yet been standardized by mainstream frameworks.
All Facts (5)
Persistent memory is typically divided into three levels: short-term working memory, episodic memory, and semantic memory
65%UnverifiedThe user profile uses a layered memory architecture divided into three layers: permanent preferences, situational physiological information and immediate conversation context.
50%UnverifiedThe solution uses a layered memory architecture: a concise "persona" stays resident in context, while a dynamic "diary" does not enter the context and is retrieved on demand via search.
50%UnverifiedLayered memory divides context into three layers: the always-loaded core state, the on-demand-loaded architecture documents and API specifications, and the background full-text-search-style history records.
50%UnverifiedMitigation solutions for long-term memory include hierarchical memory (hot storage and cold storage), TTL mechanisms borrowed from caching systems, and weight decay based on the forgetting curve; these solutions are still in an experimental stage and have not yet been standardized by mainstream frameworks.
50%