A multi-tier LLM KV-cache runtime for compressing, offloading and restoring cache in a supported memory and storage environment.
Discuss KX MemoryForgeTHE PRODUCT, IN PRACTICE
Long-context agents need retained cache, but not every deployment can keep all of it in GPU memory. KX MemoryForge addresses both cache capacity and movement: compress the supported representation, move it across qualified tiers and restore it when the workload needs it. It is an operating runtime around the cache lifecycle, distinct from a placement decision alone.
Context and simultaneous sessions can compete for the same HBM capacity.
KX MemoryForge provides supported cache compression and multi-tier retention so the team can evaluate a larger useful workload within a defined memory system.
Moving cache out of HBM is not useful if returning it introduces unacceptable delay or disrupts the session.
The runtime addresses offload and restoration together. Qualification measures the complete cache lifecycle, including movement costs and accepted serving behavior.
A cache plan involving storage, fabrics or accelerators needs a compatible data path—not only a list of possible technologies.
The delivered runtime scope specifies the supported tiers and interconnects, allowing the team to evaluate an actual offload/restore configuration.
The product scope below describes the capability and outputs to confirm for your release. Custom integration, operated delivery and additional engineering are separately agreed.
The supported compression, offload and restoration functions.
The selected memory/storage connections, which may include the documented HBM, CXL or NVMe paths.
The supported model and runtime environment in which cache movement is evaluated.
Defined capacity, quality and movement criteria for the deployed configuration; engineering work is separately scoped.
KX MemoryForge is fully available with CLI, MCP and headless access. Use the interface that fits your terminal, your agents or your automated workflow.
Run the product from a terminal or script, alongside the tools your team already uses.
Get CLI accessConnect MCP-compatible clients and agents to the product’s exposed capabilities.
Get MCP accessOperate without a graphical interface in unattended, scheduled or pipeline-driven work.
Set up headless accessProduct-specific inputs, hardware requirements and result guarantees still apply. An access mode does not change your license or deployment scope.
Explore all product interfacesA team is exploring an additional memory/storage tier before purchasing more GPU capacity.
Illustrative use of the supported scope, not a customer case study or a measured result.Choose a supported model, context workload and available tier configuration.
Exercise compression, offload and restoration in that environment.
Compare useful concurrency with restore latency, traffic and quality—not capacity alone.
A product workflow illustration, not a live execution or deployment test. Choose a stage to explore its role.
Apply a supported cache compression configuration.
Inference-platform and long-context-agent teams evaluating cache capacity across GPU memory and additional tiers.
CLI, MCP and headless access are all available. Run from a terminal or script, connect an MCP-compatible client, or use the product without a graphical interface. Choose the mode during onboarding and receive its setup instructions. Runtime requirements and supported operations remain product-specific.
The reported 4.13× sequence-capacity result belongs to the supplied 4-bit, 16K-context campaign configuration. The broader 4–10× aim is not a universal deployment guarantee. SpinMem and SpinCache remain separate products, and not every listed tier is available on every system.
The agreed KX MemoryForge release, its supported interfaces and outputs, the applicable checks and the support stated in your order. The plan defines users, workflow, environment and capacity. Source-code transfer, unrestricted redistribution and bespoke engineering are not automatically included.
Request a scoped engineering or operated-service proposal. It identifies what our team delivers, where the workload runs, infrastructure costs, acceptance requirements and ongoing support. Your software subscription alone does not activate a hosted runtime.
Explore engineeringP60 / Company-reported. The supplied campaign reports 4.13× more sequences at 16K context using its stated 4-bit configuration. The README summary reports 74 tests passing and 14 of 16 benchmarks passing. Capacity, latency and model quality remain separate measures.
These descriptions come from the supplied product and campaign material. Website checks are not an independent rerun of the engine. Qualification identifies the actual release, workload and available artifacts.
Read the evidence summaryCustomer-hosted access and a service operated by SpinForge are different scopes. Confirm the release, deployment and included support before activation.
Plans & deliveryBring a representative workload and the result your team needs. We’ll define the product, deployment and engineering scope around that task.
Activate this capability inside a qualified ai serving workflow. Compatible methods and required shared checking are included once within that workflow; the product remains independently purchasable.
Capacity, release, environment, result contract and operating responsibilities are specified before activation. Choosing an additional method is not permission to add an unrelated workload or unqualified deployment.
Explore AI Serving Choose workflow capacityCommission the model, integration and acceptance work needed for your task. Existing products stay independently available.
Use the software directly or ask us to scope the implementation.
One capability. A connected workflow.
A team to put it into practice.