The clearest thread running through the last stretch of releases is that the boring half of RAG is becoming somebody else's product. AWS moved [Amazon Bedrock Managed Knowledge Base](https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-bedrock-managed-knowledge-base/) to general availability on June 17, positioned as a fully managed RAG layer rather than an assembly of OpenSearch, chunkers, and embedding jobs you wire yourself — and the developer writeups through July have all converged on the same question of managed versus self-managed versus DIY. On the interconnect side, the Model Context Protocol shipped a dated spec revision on [2026-07-28](https://blog.modelcontextprotocol.io/posts/2026-07-28/), with vendors publishing enterprise readings of it almost immediately. Ingest, chunk, embed, hybrid retrieve, rerank: that pipeline is turning into a purchased component behind a protocol handshake.
If you have been treating your chunking strategy and reranker stack as a moat, that moat is thinning. Not gone — domain-specific parsing of contracts, EHRs, or CAD metadata still beats generic ingestion — but the default path for a new internal assistant no longer runs through a vector DB you operate.
The write path is where the work moved
The research and tooling side is pointing somewhere else. mem0 published a *State of AI Agent Memory 2026* benchmarks and trends report in mid-August, and recent work like *Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation* ([code](https://github.com/HU-xiaobai/xMemory)) argues the framing directly: an agent's memory is not a corpus you search, it's state you maintain.
That distinction has concrete engineering consequences. A document store is append-only and idempotent; you re-index and move on. Agent memory has to resolve contradictions ("the customer moved to the enterprise plan" invalidates a prior fact), carry provenance and valid-time separately from ingest time, decay or forget, and enforce tenancy at write time rather than filtering at query time.
The core tradeoff is when you extract. Write-time extraction — distilling each turn into entities, decisions, and preferences — gives you cheap reads and a queryable shape, but you pay tokens on every turn and you freeze a schema you chose before you understood the workload. Read-time extraction over raw transcripts keeps flexibility and pays latency, and re-derives the same facts endlessly. Most production systems end up hybrid, and the thing that makes the hybrid survivable is storing derived facts with a pointer back to the source span, so you can re-derive when you change models or discover your extractor was wrong.
Practical split: let the managed service own documents, which are stable and where operational toil is real. Keep derived state — profiles, commitments, resolved entities — in a store you version and can audit, with per-fact provenance and explicit invalidation. Also budget for evaluation you own regardless of vendor: if you cannot reproduce retrieval quality outside the managed KB, you cannot tell whether a regression is yours or theirs, and you cannot leave.