Subconscious is an AI lab solving the context window.
We research inference and post-training for long-horizon agents.
Inference
Codename OrangeLine
A runtime built around long agent traces, so agents run longer on the same GPUs.
Read about OrangeLinePost-training
Codename RedLine
Compression at training time, producing models that are optimized to work with it.
Read about RedLineInference
Codename OrangeLine
General-purpose inference systems waste compute on agent workloads. OrangeLine is our runtime built around agent traces, from how a request is routed to how its context is cached and compressed.
Global load balancer
Routing
Routes requests across our geographic clusters.
Cache-aware routing
Routing
Runs on our fork of NVIDIA Dynamo, optimized for prefix and suffix caching, so each step of a trace lands where its context is already cached.
OrangeLine
Inference engine
Our inference runtime. It manages context on the GPU while a trace runs: compressing the KV cache, reusing prefixes and suffixes, and freeing memory mid-run, so agents reason further and faster on the same hardware.
L3 cache
Caching
A novel L3 caching layer designed around long-horizon traces as the primary workload, so long runs keep hitting cache deep into a trace.
Subconscious Cache
Our system for KV cache pruning with preserved suffix states. When a span of context is pruned, the information it contributed survives in the retained suffix states instead of being thrown away and re-encoded.
On long tasks vs. standard inference:
- Faster task completion
2x
- Concurrent workloads
2.3×
- Max cost savings
80%
Standard runtime
Prefix cache only
Everything after the pruned span is re-encoded from scratch.
System + tools
Pruned
Recent steps
Output
OrangeLine
Prefix + suffix cache
The suffix is reused too. Some of the pruned token information is preserved in the suffix.
System + tools
Pruned
Recent steps
Output
Auto Compaction
Our runtime enables models to prune their own message list at inference time. Compaction happens automatically inside the familiar chat completions and messages APIs, and the Subconscious Cache preserves memory across those prunes.
Read the postOrangeLine in Production
OrangeLine runs in production today. Customers use it through our API and on their own GPU clusters.
Read the postLonger runs and more concurrency on the same GPUs.
The runtime manages memory and context during long agent runs.
The gains land exactly where agentic workloads hurt the most.
Handle millions of tokens with context management at runtime.
With the highly efficient Subconscious Cache, cut cost 50% and up to 80%.
Run 2.3x as many workloads on the same compute footprint.
2x faster task completion down deep reasoning chains.
SGLang
Limited to the model context window with compaction. Long tasks run with unnecessary context or compact necessary information.
Subconscious
Processes millions of tokens per run. The system compresses its context at runtime, so accuracy stays high at any scale.
Post-training
Codename RedLine
Our experimental post-training work, built on OrangeLine.
Compression at training time
RedLine compresses context during training, so long reasoning trajectories train on significantly less compute.
Models that understand compression
The models it produces are optimized to work with compressed context and suffix reuse, so they get more out of OrangeLine.
Stronger small and mid-sized models
It can significantly improve the accuracy of small and mid-sized models on unique data, tooling, and tasks.
We use it to run SFT, OPD, and RL on open-weight models for long-horizon, tool-heavy work like coding agents, browser control, and research. We also post-train models for customers on their own data and tooling.
Contact us to collaborate
Our research is informed by our connections to the academic community and our commitment to publishing our findings. Today we collaborate with labs at MIT, Princeton, Harvard, Carnegie Mellon, Tel Aviv, and Berkeley.
For research collaborations, academic partnerships, or questions about our work, reach out to hongyin@subconscious.dev.