context-engineering-advisor

Context stuffing overloads AI workflows with excessive data, degrading model attention and causing context rot. You diagnose it by identifying symptoms of degraded attention allocation and mapping out ownership of your current context boundaries to spot unbounded data injection.

Design a two-layer memory architecture by separating short-term conversational memory from long-term persistent memory. This bounded context approach prevents context rot in multi-agent workflows by governing what information persists versus what is retrieved on demand.

You need a Context Manifest for RAG deployments when retrieval strategies become unbounded and cause context rot. It defines strict boundaries to govern what information persists versus what is retrieved, ensuring proper episodic retrieval and attention allocation.

The best way to prevent context rot is applying a Research→Plan→Reset→Implement cycle. This structured cycle compresses research into a high-density plan, resets the context, and starts fresh implementation with optimized attention allocation for multi-agent workflows.

Bounded contexts work with existing RAG by defining strict ownership boundaries and episodic retrieval strategies. This structured approach limits context stuffing and ensures that only high-density, relevant information is retrieved rather than persisting everything.