What happened
KindlingAI released gr8 v1.0.1 on November 12, 2026, as a targeted patch release on top of v1.0. The release notes describe changes to the SSD-backed KV session store, the loader’s QO_CACHE writes, and project documentation and recipes. The highlighted serving change is that steered chats can now be journaled and restored across orderly restarts when the SSD store is configured with persist=on and keep=sessions or keep=all. This includes deployments using STEERING_FORCE=1.
The store extends a continuing session in place by writing newly added page blocks instead of rewriting the full conversation. The release notes say the extension reaches 1.00x write amplification for continuing steered sessions. Each stored entry carries an exact 70-byte steering profile/scale namespace. That namespace keeps an entry invisible to unsteered requests (scale=0) and to requests using other steering selections; keep=system prefix indexes remain stock-only. A continuing hit on its own resident pool also avoids redundant copies on intermediate turns.
The release reports CPU-pool tests using 64-row pages, 4,096-token reuse grids, and 3,506,176-byte page blocks. In one-chat, ten-turn tests with eviction between turns, recorded writes fell from 3,519 blocks (12.34 GB) to 512 blocks, with write amplification dropping from 6.34x to 1.00x. In a two-chat test, eight turns each while evicting one another, writes fell from 6,678 blocks (23.41 GB) to 1,152 blocks, and amplification went from 5.57x to 1.00x. An undersized-store test still had 3.85x amplification after the change, versus 4.09x before.
A live test used GLM-5.3 across four GB10 systems, with tensor parallelism set to four and a 16 GB SSD store. For a 19,000-token chat, reported first-reply latency after restart fell from 12.75 seconds cold to 3.04 seconds from SSD for plain chats, and from 12.99 seconds to 2.99 seconds for steered chats. The release says 16,384 cached tokens were restored in about 56 milliseconds, and replies were byte-identical to cold prefill for both plain and steered chats. Startup behavior also changes: on Linux, the store reserves its configured size= capacity with fallocate before opening. This prevents other files from claiming the space, but insufficient disk capacity or a filesystem without fallocate support now causes an immediate startup error.
What the archive adds
New Runtime’s earlier coverage on gr8 steering profiles described contrastive profiles and selection of up to eight profiles per request. This release adds a new fact to that story: steered sessions can now persist in the SSD store while remaining isolated by profile and scale. It is a follow-on runtime capability, not a restatement of the earlier steering feature.
Trend signal
This release supports the idea that the production boundary for an agent or model-serving system is its surrounding runtime, not just model choice. Persistent state, cache behavior, and isolation rules determine whether a capability can be used safely across requests and restarts. This is a concrete fit for harness architecture outlives model choice, though the release does not establish a broader industry trend by itself.
Why it matters
For teams using steering to produce different behavior by request, persistence is useful only if cached state cannot bleed between profiles. gr8’s namespace rule addresses that operational constraint directly, while restoring cached tokens after an orderly restart can reduce the delay before a long conversation resumes. The reported latency and write-amplification figures make the potential benefit legible for deployments with recurring sessions and SSD-backed KV storage.
The change also shifts one failure mode earlier. Previously, a cache budget larger than available space could fail later during serving; v1.0.1 reserves the configured space at boot and fails immediately if it cannot. Operators should therefore treat size= as a real disk commitment, verify filesystem support, and account for it in deployment capacity planning. The results apply to gr8’s tested configurations; they do not establish performance across other models, hardware, filesystems, or workloads.
What happens next
- gr8 operators will test persistent steered-session recovery in staging before wider production use. (3 months, 65%). Watch for: deployment notes or issues covering restart recovery, profile isolation, and SSD capacity planning.
- The release’s 1.00x write-amplification result will be tested against broader concurrency and capacity conditions. (6 months, 55%). Watch for: benchmarks beyond the stated CPU-pool scenarios and GLM-5.3 configuration.
Limits and uncertainty
The performance figures are reported by KindlingAI and have not been independently reproduced by New Runtime. The live test covers one stated model and hardware configuration, while the CPU-pool measurements cover the listed synthetic session patterns. The release does not quantify behavior under abrupt power loss or unclean shutdown; its persistence claim is specifically for orderly restarts. It also does not report the storage overhead of namespace metadata or compare durability and latency across filesystem types. Operators should validate their own workload, disk capacity, and restart behavior before relying on these results.
Skill to build now
If you operate gr8 with steered conversations, test cache isolation before enabling persistence broadly: run otherwise identical requests at two steering profiles and at scale=0, then confirm that resumed sessions do not cross those boundaries. Also stage a restart with persist=on and verify that the expected session resumes. Before deployment, set size= below confirmed free space on a filesystem that supports fallocate; the new reservation makes that capacity check a startup requirement.
Reader takeaway
The useful change is not simply that gr8 remembers more: it can persist steered sessions while keeping cache entries separated by profile and scale. Its reported write reductions and restart-latency improvements are promising for the tested setup, but teams should validate isolation and reserve enough supported disk space before putting the behavior into service.