Memindex

September 18, 2026

MCP Went Stateless and Your Connectors Have to Follow

The 2026-07-28 spec deletes sessions from the transport, changing how retrieval connectors handle pagination, tenancy, and scale.

If you maintain an MCP server that fronts a search index, a document store, or an internal wiki, the migration work is now concrete rather than theoretical. The 2026-07-28 revision removes protocol-level sessions and the `Mcp-Session-Id` header from the Streamable HTTP transport, and removes the `initialize`/`notifications/initialized` handshake entirely. Protocol version, client info, and client capabilities that were exchanged once at connection setup now travel in a `_meta` field inline on every request.

The motivation was operational. Running remote MCP servers previously required shared Redis session stores or gateway-level packet inspection, adding latency and operational cost. Google says it needed MCP to scale across millions of concurrent queries and co-founded an MCP Transports Working Group with Hugging Face and other partners to decouple the protocol from stateful transport. Anthropic frames it as a move from a bidirectional stateful protocol to a request/response model, which lets servers deploy on serverless and edge infrastructure.

What breaks in a retrieval connector

Two details matter more than the headline. First, list endpoints — `tools/list`, `resources/list`, `prompts/list` — no longer vary per-connection . If your knowledge connector shaped its tool or resource list around who was connected (per-tenant corpora, per-role document scopes), that pattern is gone from the protocol layer and has to move into arguments and authorization checks.

Second, servers that need cross-call state must now use explicit, server-minted handles passed as ordinary tool arguments (SEP-2567) . That is the right primitive for retrieval anyway: pagination cursors, result-set IDs, and expanded-context tokens become first-class, inspectable, and cacheable objects instead of implicit connection state. It also means you own their expiry and their authorization semantics. A handle that leaks across tenants is now your bug, not the transport's.

Auth tightened alongside. Six SEPs align MCP's authorization spec more closely with production OAuth 2.0 and OpenID Connect deployments , and Dynamic Client Registration is formally deprecated in favor of CIMD, continuing to work for backward compatibility but slated for removal.

Timelines and the upgrade path

Nothing forces a same-week rewrite. Upgrading is opt-in, and on Bedrock AgentCore Gateway, adding 2026-07-28 alongside 2025-11-25 does not change behavior for existing clients, since clients select a version per request. The deprecation clock is the thing to calendar: Roots, Sampling, and Logging are deprecated but supported for at least twelve months, and the legacy HTTP+SSE transport is officially deprecated with a year-long offramp. Change notifications also move to a single `subscriptions/listen` stream that clients opt into per notification type — relevant if you push index-freshness events to agents.

Who should care: anyone operating more than a handful of MCP-exposed knowledge sources behind a load balancer. Anthropic reports MCP passed 400M monthly SDK downloads, a 4x increase this year , which is roughly the scale at which sticky sessions stop being survivable.

Sources

  1. [1] Agent Memory Is Not RAG: A 2026 Production Field Guide - DEV Community
  2. [2] [2602.02007] Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
  3. [3] Retrieval-Augmented Generation for Natural Language Processing: A Survey
  4. [4] A-MEM: Agentic Memory for LLM Agents
  5. [5] Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval
  6. [6] Knowledge and Memory Beyond RAG: Why 2026 Agents Need a Write Path, Not Just a Retriever | by Micheal Lanham | Apr, 2026 | Medium
  7. [7] The Agent Memory Wars Are Here - AgentConn Blog
  8. [8] MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation
  9. [9] Agent Memory Vs RAG: What Breaks At Scale 2026 (Analyzed)
  10. [10] Anthropic’s Model Context Protocol receives anniversary update - SDxCentral
  11. [11] The 2026-07-28 Specification | Model Context Protocol Blog
  12. [12] Model Context Protocol Blog
  13. [13] Posts | Model Context Protocol Blog
  14. [14] The New MCP Roadmap | Model Context Protocol Blog
  15. [15] The 2026-07-28 MCP Specification Release Candidate | Model Context Protocol Blog
  16. [16] Model Context Protocol prepares to break with its stateful past
  17. [17] Model Context Protocol is going stateless to make scaling simpler | InfoWorld
  18. [18] Model Context Protocol
  19. [19] Guide to Enterprise Intelligence Systems 2026 | Bloomfire
  20. [20] The AI Enterprise Search Guide for IT and Knowledge Leaders
  21. [21] Enterprise Knowledge Base for AI: Architecture Guide
  22. [22] Introducing Amazon Bedrock Managed Knowledge Base for faster, more accurate enterprise AI applications | Amazon Web Services
  23. [23] Build enterprise search for agents with Amazon Bedrock Managed Knowledge Base | Artificial Intelligence
  24. [24] Enterprise Knowledge Base Software with AI-Powered Search
  25. [25] Knowledge Connectors | 50+ Enterprise Data Connectors | Katonic AI
  26. [26] AI Enterprise Search & Knowledge Management | Pinnasys
  27. [27] Enterprise Knowledge Base Platforms: Top 10 Picks for 2026
  28. [28] Dewey Long Context Embedding Model: A Technical Report
  29. [29] Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT
  30. [30] ICLR Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT
  31. [31] LongBench Pro: A More Realistic and Comprehensive Bilingual Long-Context Evaluation Benchmark
  32. [32] GitHub - gmlwns2000/LongBench-hip: LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding · GitHub
  33. [33] Long Context Benchmarks: All Three Hit 1M — Now What?
  34. [34] Evaluating Long Context (Reasoning) Ability | wh
  35. [35] GitHub - THUDM/LongBench: LongBench v2 and LongBench (ACL 25'&24') · GitHub
  36. [36] export.arxiv.org
  37. [37] TiDB
  38. [38] Top 9 Vector Databases as of September 2026 | Shakudo Blog
  39. [39] TerminusDB
  40. [40] Milvus (vector database)
  41. [41] Actian Vector
  42. [42] Top 10 Vector Databases for LLM Applications in 2026 | Second Talent
  43. [43] Chroma (vector database)
  44. [44] The 2nd Workshop on Vector Databases (VecDB)
  45. [45] TypeDB
  46. [46] The 2026-07-28 MCP Specification: A Stateless, Extensible Future | MCP Servers
  47. [47] Key Changes - Model Context Protocol
  48. [48] How AgentCore Gateway supports the MCP 2026-07-28 spec | Artificial Intelligence
  49. [49] The biggest MCP spec update ships July 28: What changes for AI agent authentication — WorkOS
  50. [50] Scaling AI Agent Infrastructure with the MCP Stateless updates - Google Developers Blog
  51. [51] MCP 2026-07-28 spec: stateless core, coming to Claude | Claude by Anthropic
  52. [52] MCP 2026-07-28: From Local Tool to Distributed Protocol - Agentic AI Foundation (AAIF)
  53. [53] 'Model fatigue' sets in AI labs roll out new versions at ...
  54. [54] AI Model Release Tracker | Evertune
  55. [55] New Models Today — AI & LLM Releases Last 24 Hours | Price Per Token
  56. [56] Vertex AI release notes | Google Cloud Documentation
  57. [57] Anthropic, OpenAI, Google and Meta All Update AI Models This Week - AI Agent Blog | AgentLocker
  58. [58] AI Updates Today (September 2026) – Latest AI Model Releases
  59. [59] Anthropic Release Notes - September 2026 Latest Updates - Releasebot
  60. [60] AI Week in Review 25.03.15
  61. [61] the sequence radar 803 last week
  62. [62] AI Agents News — Week of September 13, 2026 (Daily Updates)
  63. [63] Everyone Is Building Agent Memory Now: What the Wave Proves - DEV Community
  64. [64] AI Agent Memory 2026: Vector, Graph, Episodic Update
  65. [65] Whoever Owns the Memory Owns the Agent | by Baheet | Aug, 2026 | Medium
  66. [66] The 20 Best AI Agent Memory and Context Tools for P… | StartupHub.ai
  67. [67] TencentDB Agent Memory Tops 20,000 GitHub Stars in 90 Days, Launches Team Memory for Multi-Agent Collaboration
  68. [68] State of AI Agent Memory 2026: Benchmarks & Trends Report
  69. [69] Agents that remember: introducing Agent Memory | Cloudflare Blog
  70. [70] Agentic AI News — September 2026 Launches, Models & Research | Agentic.ai

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.

Memindex — MCP Went Stateless and Your Connectors Have to Follow