Two things worth reading landed in the past week or so, and they rhyme. Mem0 published a *State of AI Agent Memory 2026* survey of benchmarks and trends. And the Model Context Protocol project posted a refreshed roadmap (blog post dated August 23, site roadmap updated August 22), the first major direction-setting since the 2026-07-28 specification. Separately, the weekly Anthropic release cadence continues to be dominated by plumbing — a Claude ecosystem roundup on September 21 led with Messages API beta changes, MCP growth, and SDK updates, and there were fresh Claude Code releases within the last two days.
None of that is a model launch, which is the point. The interesting work right now is below the model.
The read path was never the hard part
Most teams that shipped a RAG system in the last two years built a read path: chunk, embed, retrieve, rerank, stuff. That pipeline is now close to commoditized — Bedrock's managed knowledge base and ChatGPT's company-knowledge connectors both exist precisely because the read path is generic enough to sell as a service.
What isn't commoditized is the write path. When an agent finishes a task, something has to decide what was learned, whether it contradicts an existing memory, whose memory it belongs to, and when it expires. That's extraction, dedup, conflict resolution, decay, and provenance — five decisions your vector store does not make for you. Retrieval quality metrics don't measure any of them. Recall@k looks fine while the store quietly accumulates two contradictory facts about the same customer and returns whichever one embeds closer.
If you're evaluating agent memory this quarter, instrument the write side: contradiction rate per 1k writes, median age of retrieved facts, share of retrievals traceable to a source document. Those tell you more than another reranker bake-off.
Why the MCP roadmap matters here
MCP's push toward a leaner, more stateless server model pushes state ownership back to the host application. That's the right layering — a connector shouldn't be a database — but it means the memory store becomes your problem, not your tool provider's. Teams that assumed MCP servers would carry session context will need an explicit store with its own access control.
That cuts in a specific direction for buyers. If you adopt a managed knowledge base, you inherit its chunking, its refresh semantics, and — most consequentially — its ACL model. Permission-aware retrieval is where managed enterprise search either saves you a year or quietly leaks a comp spreadsheet. Ask vendors how deletions and permission revocations propagate to embeddings, and how long the window is. Few answer cleanly.
The takeaway for builders: budget engineering for writes and forgetting, not just for recall.