The most useful number this week came from a memory vendor, not a model lab. Micron said AI context is growing roughly 30x a year, and laid out a roadmap with HBM6 arriving by 2030 ([FourWeekMBA summary, Oct 22](https://fourweekmba.com/ai-micron-says-ai-context-grows-30x-a-year-plans-hbm6-by-2030/)). Hold those two figures next to each other. Context demand is compounding on an annual cycle; the bandwidth and capacity to serve it is on a multi-year silicon cycle. That gap is the whole argument for retrieval, and it is not closing.
If you have been waiting for context windows to get long enough that you can delete your retrieval layer, this is the clearest signal yet that the wait doesn't end. The constraint was never the advertised token count. It's KV cache: attention state grows with sequence length, it lives in high-bandwidth memory, and it is per-request. A 10M-token window you can technically fill is a window you cannot afford to fill concurrently across a thousand users. Long context is a capability; it is also a per-request bill denominated in a resource that improves on a hardware cadence.
The practical consequence for anyone building knowledge tools: the question is not "retrieval or long context," it's what you're willing to pay to keep resident. Treat the window as a cache tier, not storage. Things that earn residency: the current task state, the last few turns, the handful of documents the agent is actively editing. Things that belong in retrieval: everything that might be relevant. Things that belong in a compaction step: the long tail of turns that matter as a summary and nothing more. This is ordinary memory-hierarchy design, and it's why agent memory systems and RAG pipelines keep converging on the same shape — a hot window, a warm summary store, a cold index — regardless of which framework they started in.
Two smaller signals point the same direction. MemoraX said it had 10 papers accepted to NeurIPS 2026 ([Oct 22](https://www.globenewswire.com/news-release/2026/10/22/3386687/0/en/memorax-ai-researchers-and-collaborators-have-10-papers-accepted-to-neurips-2026.html)), which tells you memory is now a research track with institutional backing rather than a framework feature. And Vectorial closed a $3M seed ([FINSMES, Oct 22](https://www.finsmes.com/2026/10/vectorial-raises-3m-in-seed-funding.html)) — small, but seed money still flowing into the retrieval substrate is not what you'd see if long context had eaten the category.
Meanwhile the connector layer keeps consolidating around MCP: Bloomberg shipped an enterprise MCP offering in late September, and Atlassian and OpenAI expanded their partnership around enterprise knowledge earlier this month. Those move where your data comes from. They don't change what fits in the window.
Who should care: anyone sizing a 2027 architecture. Budget for retrieval and compaction as permanent components. The hardware roadmap says you have four years before the tradeoff meaningfully shifts, and your users will have raised their expectations several times over by then.