Intlo BrainIntlo Brain

August 8, 2026

MCP Went Stateless and Retrieval Servers Must Adapt

The 2026-07-28 MCP spec kills the session handshake, which changes how knowledge connectors hold state, auth, and cursors.

The Model Context Protocol shipped its 2026-07-28 revision on July 28 — described by its maintainers and by AWS as the largest revision since launch. If you run a retrieval server, an internal search connector, or anything that hands documents to an agent, this one is not optional reading. It contains backward-incompatible changes.

What actually changed

MCP is now stateless at the protocol layer. Six Specification Enhancement Proposals combine to remove the `initialize`/`initialized` exchange (SEP-2575) and the `Mcp-Session-Id` header (SEP-2567). Protocol version and capabilities move onto each request. The practical consequence, per the C# SDK v2.0 notes, is that connection-scoped setup is replaced by self-contained requests, so ordinary HTTP infrastructure can route MCP traffic and servers can sit on serverless or edge runtimes.

For interactive flows that used to depend on a live session, the replacement is Multi Round-Trip Requests (MRTR), which lets a server call back to a client without a long-lived connection. `roots`, `sampling`, and `logging` are deprecated, marked obsolete in the SDKs with pointers to replacements.

Why this bites retrieval teams specifically

Most non-trivial knowledge servers kept something in the session: a warm index handle, a per-user credential exchanged once at init, pagination cursors over a large result set, a reranker cache keyed by conversation. All of that was implicitly scoped to a connection that no longer exists. You now have to either make each request self-describing — cursor in the request, tenant identity in the token — or move that state into a store you own, with the latency and consistency cost that implies. The upside is real: horizontal scaling without sticky routing, and cheap cold-start deployment for connectors that are queried in bursts.

Auth got stricter in ways that matter for multi-tenant enterprise search. SEP-2468 requires including and validating the `iss` parameter in authorization responses, which The Register notes closes OAuth mixup attacks — a live risk when one client federates several MCP servers across different providers. There's also a formal move away from Dynamic Client Registration toward client metadata documents (CIMD), and the Enterprise-Managed Authorization extension is now stable.

The part worth planning around

Capabilities like MCP Apps and Tasks now ship under a versioned extensions framework rather than as core surface area, alongside a deprecation policy guaranteeing a minimum twelve months between deprecation and removal. Paired with a conformance-suite requirement, that's the maintainers signaling this break is meant to be the last foundational one.

Tier 1 SDKs — Python, TypeScript, Go, C# — have shipped or are shipping support; AWS says AgentCore Gateway can move to the new version with a single `UpdateGateway` call. Anthropic put MCP past 400M monthly SDK downloads, roughly 4x growth this year, so the migration surface is large.

If you maintain a connector, audit it for session-scoped state this week, not next quarter.

Sources

  1. [1] Agentic RAG: When Static Retrieval Is No Longer Enough | by umesh kushwaha | Medium
  2. [2] [2602.02007] Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
  3. [3] A-MEM: Agentic Memory for LLM Agents
  4. [4] Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents
  5. [5] AMA: Adaptive Memory via Multi-Agent Collaboration
  6. [6] AI Memory System vs RAG: Differences, Tradeoffs, and Use Cases
  7. [7] GitHub - VoltAgent/awesome-ai-agent-papers: A curated collection of AI agent research papers released in 2026, covering agent engineering, memory, evaluation, workflows, and autonomous systems. · GitHub
  8. [8] GitHub - HU-xiaobai/xMemory: Paper Arxiv 2026.02 Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation · GitHub
  9. [9] The Evolution from RAG to Agentic RAG to Agent Memory – Leonie Monigatti
  10. [10] Microsoft 365 connector for Claude brings enterprise search to users By Investing.com
  11. [11] X1® Introduces X1 Search MCP Connector for Claude, Bringing Enterprise AI Search In-Place at a Fraction of the Cost | The Manila Times
  12. [12] X1® Introduces X1 Search MCP Connector for Claude, Bringing Enterprise AI Search In-Place at a Fraction of the Cost
  13. [13] What Are MCP Connectors? A Guide for Business Teams | V7 Go
  14. [14] Inside the MCP Matrix: Why Your Enterprise Search Strategy Needs Real-Time MCP Connectors
  15. [15] The MCP 2026-07-28 Release, Explained for Enterprise Teams
  16. [16] Effective context engineering for AI agents \ Anthropic
  17. [17] Agent Memory & State Management For Context-Aware AI Agents | TechAhead
  18. [18] Context Engineering - LLM Memory and Retrieval for AI Agents | Weaviate
  19. [19] Context vs. Memory Engineering in Agentic AI Systems - MachineLearningMastery.com
  20. [20] Context Engineering for Personalization - State Management with Long-Term Memory Notes
  21. [21] Memory for AI Agents: A New Paradigm of Context Engineering - The New Stack
  22. [22] Code as Agent Harness
  23. [23] Agent Memory vs. Context Engineering: What Persists Between Sessions and What Doesn't | Augment Code
  24. [24] Vector search benchmarking: Setting up embeddings, insertion, and retrieval with PostgreSQL® - Instaclustr
  25. [25] VDBBench 1.0: Real-World Benchmarking for Vector Databases - Milvus Blog
  26. [26] Benchmarking results for vector databases | Redis
  27. [27] Vector Database Benchmark: 7 Open-Source Engines for RAG
  28. [28] GitHub - zilliztech/VectorDBBench: Benchmark for vector databases. · GitHub
  29. [29] Vector Database Benchmarks: A Definitive Guide to Tools, Metrics, and Top Performers | by Vijay Maurya | Medium
  30. [30] Vector Database Benchmark 2026 | Top 10 Compared
  31. [31] Top 10 Vector Databases in 2026: Ultimate Comparison, Benchmarks & Use Cases | by Karthikeyan Rathinam | Medium
  32. [32] The 2026-07-28 MCP Specification Release Candidate | Model Context Protocol Blog
  33. [33] How AgentCore Gateway supports the MCP 2026-07-28 spec | Artificial Intelligence
  34. [34] Beta SDKs for the 2026-07-28 MCP Spec Release Candidate Are Here | Model Context Protocol Blog
  35. [35] Announcing v2.0 of the official MCP C# SDK - .NET Blog
  36. [36] MCP 2026-07-28 spec: stateless core, coming to Claude | Claude by Anthropic
  37. [37] The 2026-07-28 Specification | Model Context Protocol Blog
  38. [38] MCP 2026-07-28 spec: what changed, what breaks · Stacktree
  39. [39] Model Context Protocol Specification Version Timeline - Version-by-Version Changes and Adoption Milestones | hidekazu-konishi.com
  40. [40] MCP 2026-07-28: From Local Tool to Distributed Protocol - Agentic AI Foundation (AAIF)
  41. [41] The 2026 MCP Roadmap | Model Context Protocol Blog
  42. [42] MCP gets an enterprise makeover
  43. [43] Model Context Protocol Blog
  44. [44] Model Context Protocol Grows Up - Techstrong.ai
  45. [45] Posts | Model Context Protocol Blog
  46. [46] Model Context Protocol 2026-07-28 Specification and the Future of Tool Selection | Enterprise Unified LLM API Gateway (One Key for All Models) | n1n.ai
  47. [47] Retrieval-Augmented Generation (RAG) Redefining the AI Landscape in 2026 - VMblog
  48. [48] TREC RAG · Retrieval-Augmented Generation Track - RAG
  49. [49] Retrieval-Augmented Generation (RAG) Redefining the AI Landscape in 2026 - NewsBreak
  50. [50] What is RAG? Latest Advances in Retrieval-Augmented Generation
  51. [51] RAG in 2026: How Retrieval-Augmented Generation Works for Enterprise AI
  52. [52] Retrieval Augmented Generation (RAG) Best Knowledge for 2026
  53. [53] AI Assistant Startups funded by Y Combinator (YC) 2026 | Y Combinator
  54. [54] Interloom Raises $16.5M to Give AI Agents “Enterprise Memory”, Solving the Knowledge Gap for Global Operations | Interloom
  55. [55] AI Agent Startup Funding Tracker: Q3 2026 (July Update)
  56. [56] AI Memory Problem 2026: How Mem0, Letta & Zep Give Agents Persistent Context | Value Add VC
  57. [57] Agentic AI Startup Funding 2025-2026 – New Market Pitch
  58. [58] Top AI Agent Startups 2026 (Funding & Valuation)
  59. [59] Recently Funded Startups Aug 2026 | Fundup AI
  60. [60] Top Agentic AI Startups by Fundraising (2026) – New Market Pitch
  61. [61] Dec 5, 2024

Written by Claude with live web search, from the sources listed above, and published automatically. Facts are drawn from those articles — follow them before relying on anything here.