Two-system split
A small high-attention reasoner plus an unbounded memory runtime — unlimited external state, a small working set in attention.
← Store · // product
subnext
subnext is a Dynamic Context Runtime (DCR): a zero-dependency Rust engine that pairs a small high-attention reasoner with an unbounded memory runtime, so a model works from a small planned working set instead of an ever-growing context window. Support and integration from $299.
from $299support & integration
⚠ Research-grade softwaresubnext is actively developed research-grade infrastructure. The design and benchmarks are published so you can verify them; validate behaviour against your own workload before relying on it in production. Provided as-is, without warranty.
// what it is
Unlimited history is not unlimited attention
subnext is a Dynamic Context Runtime (DCR) — a specification plus a working, zero-dependency Rust implementation. Its thesis is simple: unlimited history is not the same as unlimited attention. Recursive language models move the context-rot problem out of the transformer and into a program; subnext takes the next step and makes that program a runtime.
It splits the system in two: a small, high-attention Reasoner, and an unbounded Memory Runtime that decides which representation of the history to hand it. Incoming context is indexed into external state, and only a small, planned working set is ever placed in front of the model.
The repository ships worked examples and benchmarks — DCR against full context, sliding window, RAG, summarize-all and recursive baselines — so the behaviour is something you can run and measure, not just read about.
// capabilities
What it does
State indexer
Incoming context is indexed into external state and recalled as the right representation on demand, instead of being stuffed into the window.
Correction & recompute
Worked example covers correction, exact-quote recall, justification and recompute — the primitives a long-running agent actually needs.
Built-in benchmarks
bench compares DCR against full context, sliding window, RAG, summarize-all and recursive baselines out of the box.
Zero dependencies
Pure Rust with no external crates to audit or pin — small, inspectable, and easy to drop into an existing service.
Spec plus implementation
The design lives in docs/; the runtime lives in src/, so the behaviour and its justification travel together.
// under the hood
Specifications
- Language
- Rust (zero-dependency)
- Type
- Library + CLI (demo, bench)
- License
- MIT
- Run
cargo run --release -- demo/-- bench- Pairs with
- Our Peregrine inference engine and the DCR research write-up
- Intended use
- Long-context agents, assistants and retrieval pipelines
// what you pay for
Support, not a paywall
subnext is open source (MIT). What you pay for is getting it into your stack and tuned to your workload.
- Integration help — wired into your model serving and agent loop, with the working-set policy tuned to your data.
- Custom memory representations — indexers and recall strategies built for your documents, tools and latency budget.
- Benchmarking on your workload — DCR measured against your current context strategy on your own traffic.
- A direct line to the engineer who wrote it — no support tier in between.
// the fine print
Terms & liability
As-is, without warranty. subnext is research-grade software provided for evaluation and integration. To the maximum extent permitted by law, Cyber Sec and Stephan Botes accept no liability for any damage, data loss or consequence arising from its use.
What the fee buys. The price shown is a starting point and covers the engineer’s time — integration, custom representations, benchmarking and fixes. The source is MIT, so you are paying for the work, not for access.
Questions about scope, licensing, or authorization? Get in touch first.