Agent memory is evolving from passive retrieval-based RAG

Agent memory is evolving from passive retrieval-based RAG toward active execution-state management, where models dynamically ‘sculpt’ and reconstruct their context window to maintain coherence.

The ‘Context Sculpting’ approach suggests that model-led editing of the context window improves performance and resource efficiency over append-only methods.
Research on ‘Memory as Execution State Management’ and ‘Memory is Reconstructed, Not Retrieved’ proposes a shift from static vector search to relational, graph-based session states.
The Universal Memory Protocol (UMP) and AdMem framework aim to standardize how persistent knowledge is structured and updated across diverse agent environments.

Standard vector-similarity RAG is becoming insufficient for long-horizon agents, necessitating more structured, relational memory controllers.
Builders can reduce token costs and ‘context rot’ by implementing agents that actively manage and prune their own working memory.

medium