Vector search remains crucial in the age of LLMs, contradicting the idea that advancements in large language models have rendered it obsolete. LLMs often err not due to inherent weakness but a lack of access to necessary contextual information scattered across various sources. Vector search provides agents with a vital memory layer for inspecting, querying, and rebuilding context from disparate data. This memory capability is essential for open agentic infrastructure, ensuring agents can function across different tools and runtimes. Treating prompts as storage layers leads to unmanageable complexity, whereas vector search enables efficient retrieval of relevant project context. It allows agents to answer specific queries about code migrations, tool failures, documentation, and past decisions. Agents require both immediate working memory and long-term reference memory, which vector search facilitates by fetching pertinent information instead of overwhelming the model. The MCP interface formalizes connections between AI applications and external systems, seamlessly integrating vector search capabilities. This abstraction allows for portability, ensuring agents can operate across different environments. Agent memory should be populated with manually retrieved information, indexed with comprehensive metadata. Retrieval should be precise, providing source links and utilizing metadata filters rather than relying solely on similarity. Agents should be equipped with enough source data to verify information and avoid blind trust. Vector search, when combined with provenance, makes agent memory highly effective and debuggable. A practical agent memory loop involves chunking by meaning, embedding, and storing data with metadata, followed by runtime retrieval based on agent queries. This retrieval process, mediated by MCP, allows agents to access relevant context for decision-making. Debugging agent errors becomes manageable by isolating retrieval failures from reasoning failures. Inspectable retrieval layers enable targeted improvements to chunking, filtering, and ingestion processes. Ultimately, vector search hasn't been replaced but has become more valuable as agents gain more external resources to access.