Memory Became a Market: The AI Data Stack is Being Rebuilt on Three Fronts

Fourteen months ago this index added Memory Wall as a pain point and started mapping the memory layers forming around it. On May 28 we wrote that memory had become the attack surface. In July 2026 the other shoe dropped: memory became a market — with everything a market brings. Contested benchmarks. Funding rounds. Deprecations. And two critical CVEs in the most-adopted memory layer.

That's one front. The same month produced decisive movement on two others: the columnar file format that underpins nearly every data lake is fracturing into workload-specific successors, and the transformer KV cache — the fastest-growing dataset nobody planned for — got both a production-grade compression algorithm and a protocol for living on S3. Three fronts, one direction: the AI data infrastructure stack is being rebuilt in real time, and every rebuild lands on object storage.

Front one: the agent-memory benchmark war

The competitive shape of the memory-layer market snapped into focus this month, and the clearest evidence is that the vendors can no longer agree on what the numbers are.

Mem0's research page publishes 92.5 on LoCoMo and 94.4 on LongMemEval for its token-efficient memory algorithm. Its July "State of AI Agent Memory" blog rounds the LongMemEval claim to 94.8. A third source in the same week — a comparison page — cites 93.4 end-to-end. All vendor-published, all slightly different.1 Meanwhile a third-party whitepaper reports Letta ahead of Mem0 by 34.4% — on LoCoMo-Plus, a different benchmark variant, with run configurations that haven't been published for reproduction.2 Cognee sidestepped the crowded benchmarks entirely and published against BEAM, the 10-million-token long-horizon suite this index tracks as the successor to both LoCoMo and LongMemEval.3 And newer entrants — MemPalace, agentmemory — are posting retrieval-only Recall@5 scores above 95%, numbers that measure a different (and easier) thing than end-to-end QA accuracy.4

None of these numbers are directly comparable. That is the point. When every vendor picks the benchmark it wins, the benchmark stops being a measurement and starts being a marketing surface. We said it in the Mem0 node and it bears repeating in a post: no cross-vendor harness is yet authoritative, and any architecture decision hanging on a single vendor's benchmark table is hanging on air.

The market signals beyond the benchmarks are more legible. Supermemory raised $3M to chase the managed-memory segment.5 Mem0 formally deprecated OpenMemory — its self-hosted MCP memory server — in May, consolidating on the v2 SDK line. And the security story that our May 28 post anticipated kept compounding: after the CVSS 8.1 injection vulnerability patched in April, July brought disclosure of CVE-2026-59705 — an unauthenticated-access flaw in the OpenMemory API, CVSS 9.3, in which FastAPI routers shipped without authentication middleware and authorization trusted a client-supplied user ID. Unauthenticated attackers could read, modify, or delete arbitrary tenant memories, and a global_pause parameter offered platform-wide denial of service.6

Two critical CVEs in one memory layer in one quarter is not a Mem0 story. It is a category story: agent memory went from prototype plumbing to production attack surface faster than its engineering hardened, exactly the trajectory we mapped in May. The deployment pattern that follows is the one this index now encodes on the Mem0 node — memory layers sit behind an authenticating reverse proxy, never bare on a network, and the object storage underneath them stays sealed off from the API surface.

Front two: the file format fracture

For a decade, the answer to "what format is the data lake in" was Apache Parquet, full stop. The July discourse — a widely-shared "Parquet killers" analysis, a Dremio landscape survey, a file-format-renaissance roundup — makes clear that era is ending, but not the way the headlines frame it.7

The succession is not one format replacing another. It is the physical file layer fracturing by workload. Lance owns random-access retrieval — the pattern AI dataloaders and vector search actually use — with miniblock layouts replacing row groups entirely. Nimble, Meta's entry, owns decode velocity on ML feature tables tens of thousands of columns wide, explicitly trading storage footprint for CPU efficiency. Vortex owns the general-purpose analytical slot with cascading per-segment encodings and zero-copy Arrow-native reads. Parquet keeps the bulk warehouse scans it was designed for. Above all of them, Apache Iceberg becomes the unifying table layer that makes multi-format lakes manageable rather than chaotic.

Two caveats belong in the record. First, the flagship performance claims — Lance's random-access multiples, Vortex's 100× point reads — remain overwhelmingly vendor-published; independent production case studies are the conspicuous gap in this space, and we'll index them the week they exist. Second, Meta actively discourages third-party reimplementation of the Nimble spec to prevent fragmentation — a governance posture opposite to Parquet's ecosystem-first history, and one worth watching as a tell for how open the post-Parquet era actually is.

Front three: the KV cache moves to object storage

The transformer KV cache is the dataset nobody provisioned for. A single high-end GPU generates on the order of terabytes of it daily, most discarded after one use because HBM is finite and recomputation was assumed cheap.8 In 2026 both assumptions failed at once — and the responses arrived from opposite directions.

From the compression side: Google's TurboQuant, peer-reviewed at ICLR 2026, compresses KV tensors to 3-bit codes at runtime with no calibration or fine-tuning, and was merged into vLLM in April with production presets. What elevates it beyond another vendor claim is that it now has an independent reproduction: a community measurement on consumer AMD hardware ran a 31B model at a full 256K context in 27.1 GB versus 44.1 GB unquantized — the first non-Google confirmation of the memory numbers.9 (Honest counter-signal, also in the record: llama.cpp evaluated the port and declined it in June.)

From the tiering side: LMCache decoupled cache management into its own process, doing parallel lookups across CPU memory, NVMe, and S3. And ObjectCache — the arXiv design this index added in June — closes the loop by making S3 itself a hot tier: KV blocks stored as hash-addressed chunks, fetched layerwise so that object-store latency hides behind the previous layer's attention compute. On 100 Gbps RoCE it holds within ~5.6% of local DRAM latency at 64K contexts.10

Put the two directions together and the destination is explicit: S3-compatible object storage as a page-addressable KV appliance — ICMS tiers in the middle, quantization shrinking the bytes, layerwise protocols hiding the latency. The KV-Cache Disaggregation architecture stopped being a research direction and started being a deployment pattern.

The floor under all three fronts

Why is all of this converging on object storage now? Because the alternative got repriced out from under the industry.

The NAND Flash Supply Crisis — added to the index on July 21, 2026 as a first-class pain point — is the economic floor under every architectural story above. Enterprise SSD prices rose roughly 472% in twelve months as AI demand crowded component supply; a 30TB enterprise drive that cost ~$3,000 in mid-2025 was quoted near $11,000 by April 2026, with server DRAM on a doubling trajectory.11 Vendors passed it through: Pure Storage disclosed ~70% cumulative customer price increases in a CEO letter, and Wasabi and Backblaze raised prices in the same window.12

At those prices, "keep the growing dataset on local flash" stops being an architecture and becomes a budget incident. Agent memory archives, feature stores, training lakes, KV-cache pools — the datasets that compound daily — tier onto object storage not because it's elegant but because it's the only line in the TCO model that still scales. Flash gets rationed to the latency-critical hot path. Everything persistent lands on S3. Which is to say: the Memory Wall squeezes from above, the component market squeezes from below, and Tiered Storage on object storage is the release valve for both.

Our read

(2026-07-21 — this is the editorial register; the observable edges live on the nodes.)

The three fronts are one story. Agent memory, columnar formats, and KV-cache infrastructure are each specializing above object storage while the economics forcibly consolidate persistence onto it. We expect the next two quarters to bring: a serious attempt at a cross-vendor agent-memory harness (the benchmark chaos is unsustainable for buyers), the first independent production case studies for the post-Parquet formats, and at least one managed-cloud product that sells KV-cache-on-S3 as a line item rather than a research pattern. The full call, with revision conditions, is on the ledger.

What would make us wrong: flash prices normalizing fast enough that local-first architectures stay economical (watch fab capacity announcements), or the memory-layer market consolidating by acquisition before a neutral benchmark forces comparability.

Works cited

Footnotes

  1. Mem0 Research — vendor-published 92.5 LoCoMo / 94.4 LongMemEval; the 94.8 variant appears in Mem0 — State of AI Agent Memory 2026; the 93.4 end-to-end figure in Mem0 vs Letta comparison. ↩

  2. appliedAI — Agentic Memory Realized (PDF) — third-party report of Letta +34.4% on LoCoMo-Plus; run configs unpublished. ↩

  3. Cognee — research and evaluation results — BEAM long-horizon results; harness at MFlow benchmarks. ↩

  4. preuve.ai — AI Memory Systems Statistics 2026 — aggregated stats including MemPalace / agentmemory Recall@5 claims; retrieval-only metrics, not end-to-end QA. ↩

  5. Supermemory — raise announcement — vendor-announced seed round. ↩

  6. CVE-2026-59705 (cve.org) — unauthenticated OpenMemory API access, CVSS 9.3 per the CNA record; NVD entry. The April injection CVE: GHSA-5gv3-2fv6-jvhx. ↩

  7. Materialized View — Nimble and Lance: The Parquet Killers; Dremio — the evolving file format landscape; The File Format Renaissance (dev.to). ↩

  8. Most Tokens Should Never Be Recomputed (Medium) — the KV-generation-volume framing and LMCache decoupling story. ↩

  9. Independent AMD RDNA4 measurement (Reddit) — 27.1 GB vs 44.1 GB at 256K context; AWS ML Blog — GPUDirect + TurboQuant on FSx for Lustre — the AWS reference architecture. ↩

  10. arXiv 2605.22850 — ObjectCache: Layerwise Object-Storage Retrieval for KV Cache Reuse — layerwise S3 delivery, ~5.6% latency overhead vs DRAM at 64K on 100 Gbps RoCE. ↩

  11. DropReference — SSD price increase 2026 — supply-chain coverage of the 472% escalation; direction corroborated by vendor pricing actions. ↩

  12. Akave — the storage squeeze — cross-vendor pass-through coverage quoting the Pure Storage CEO letter (April 23, 2026). ↩