The MCP maintainers published an updated roadmap this week, and the two items most relevant to anyone building retrieval systems are not framed as retrieval at all.
The post covers the next specification release and beyond, and was developed by the Core Maintainers together with community maintainers and Working Groups. It follows a March roadmap organized around transport evolution and scalability, agent communication, governance maturation, and enterprise readiness — most of which landed in the 2026-07-28 specification release. The headline change there: protocol-level sessions and the initialization handshake are gone, so a server can scale horizontally without holding state (SEP-2575, SEP-2567). If you run connectors behind a load balancer, that already removed a sticky-session constraint from your deployment.
Tool surfaces have the same problem as chunk stores
The new roadmap names something most teams have hit in production. Connecting to a server with a hundred tools means the model pays for that entire surface before the user has asked a single question, and tool selection tends to get worse as the list grows. The response is to begin work on progressive discovery.
This is the top-k problem with a different label. A tool manifest is a corpus; the model does approximate nearest-neighbor matching over names and descriptions in-context; precision degrades as the corpus grows and near-duplicates multiply. Teams have been solving it privately with tool-retrieval layers — embed the manifest, retrieve the ten plausible tools, expose only those. Standardizing discovery at the protocol level means that layer stops being bespoke glue per host application. If you maintain a large internal MCP server, the design question shifts from "how do I trim my tool list" to "what discovery signals do I expose so a client can page through my surface intelligently."
The second item is quieter but bites harder in RAG pipelines. A tools/call response can carry the same output in more than one form, and a server developer has no way to know which form a given client will put in front of the model; the roadmap aims to standardize on one clear contract. Anyone who has shipped a retrieval tool knows the failure: you return structured JSON plus a text rendering, the client picks the one you did not optimize, and your grounding quality silently drops. Non-determinism in what actually reaches the context window makes evaluation nearly meaningless, because your offline eval and the client's real payload diverge.
The parallel in memory research
The same convergence shows up in the literature. A recent arXiv paper argues that standard RAG is poorly matched to agent memory, because agent memory forms a bounded, coherent interaction stream where many spans are near-duplicates, so flat top-k retrieval returns redundant context while summary hierarchies blur distinguishing detail. Their proposal is to decouple reusable facts and updates before aggregating them.
Tool manifests, memory streams, document corpora — three cases where flat similarity search over redundant items is the wrong default, all being rethought at once.