TierSense
Understand difficulty. Decide when to compress and recall.
The main TierSense repository: use the three modules independently or combine them for your application. Find project navigation, integration guidance, and the centralized authors' statements here.
[!IMPORTANT] Authors' statement | Independent technical approach and research timeline
We deeply regret that, in this day and age, publishing our work requires us to clarify when we began it. Nevertheless, we will do so.
We solemnly state that the three methods and two manuscripts were completed before September 14, 2026, Beijing time (UTC+8), with their actual completion dates being earlier still (supported by the WeChat screenshot at the end of this page). Jev appeared on September 15, 2026. This work represents approximately six months of effort before its public release. These were not papers hastily assembled with AI within a few days of Jev's appearance. The two papers emerged as branches of the research process; both are complete studies developed through validation and released recently after repeated refinement. For the detailed timeline, please contact the account holder directly.
This method follows a completely different technical approach from Jev and includes our own mechanistic findings. We regard it as an advance beyond Jev through an independent technical direction—not a reproduction or a rushed follow-up. This is the authors' assessment of originality and technical direction, not a claim that a universal performance advantage across all tasks has already been demonstrated in published comparisons.
We oppose hastily assembled work lacking substantive research presented under the label “auto research,” and we disagree with applying that label to carefully developed work based solely on publication timing. We want to make a genuine contribution to the community and recognize that AI-assisted research can produce high-quality results. We ask that the effort behind this method and manuscript not be dismissed as automatically generated filler without a fair examination.
At the same time, we understand the concerns about research quality and appreciate those who speak up for meaningful scholarly discussion. We hope to work alongside you to produce and publish genuinely valuable research, so that sustained effort receives thoughtful consideration rather than being dismissed overnight as filler. If this resonates with you, please join our official discussion group to exchange ideas, explore new directions, and pursue more interesting and rigorous research together.
We welcome you to use our work and share your valuable feedback. Questions, reproduction efforts, and open discussion are all welcome.
Three modules, one entry point
| Module | Question | Output | Guide |
|---|---|---|---|
| TierSense-Score | How difficult is the current task? | Five feature scores and an overall score | Difficulty scoring |
| TierSense-Compress | Which historical steps are ready for compression? | Per-step recommendations and span selection | History compression |
| TierSense-Memory-Decide | Should history be compressed or recalled now? | Two decisions and their confidence values | Memory decisions |
These modules provide decision signals. Your application switches models, generates summaries, and retrieves memories. See each module's documentation for its complete contract and latest capabilities.
Choose a module
- Difficulty display or a signal for model routing: use Score and define your own score-to-model mapping.
- Selecting history to compress: use Compress, then your own summarizer.
- Deciding when to use an existing memory system: use Memory-Decide and check execution preconditions.
- Combining modules: read the integration guide. You do not need all three APIs on every request.
Quick start
Get an API key from the TierSense platform. These endpoints follow the current module documentation; creating this hub does not change them.
| Capability | POST endpoint |
|---|---|
| Difficulty scoring | https://tierflow.cn/tiersense/v1/score |
| History compression selection | https://tierflow.cn/tiersense/v1/compress |
| Compression / recall decisions | https://tierflow.cn/tiersense/v1/memory/decide |
Minimal request examples for all three APIs. Request and response contracts differ; fields are not interchangeable.
Paper
Memory Control Signals Emerge Before Action in Long Horizon Agents
StateComp: Learning When to Compress History in Long Horizon Agents
Repository PDF copies: Memory Control · StateComp
Read or download the full author-provided PDFs: Memory Control (35 pages) and StateComp (33 pages). Research systems and hosted APIs have different scopes; see the usage guides for the API contracts.
Documentation
- Quick start: minimal requests for the three APIs.
- Agent integration: invocation points, application responsibilities, and limitations.
- Authors' statements: centralized statements and research timeline material.
This repository is the single maintained home for documentation, integration examples, and paper links across all three modules. Please direct future updates, Issues, and Stars here. The old repositories retain migration notices, historical files, and commit history. Model weights, production configuration, and self-hosted inference server code are not included.
- Request examples: select Score, Compress, or Memory-Decide from one command.
- Report an issue: the shared issue tracker.
- Migration notes: old repositories and new module directories.
Join the community
We welcome you to use our work and share your valuable feedback. Questions, reproduction efforts, and open discussion are all welcome.
Research timeline screenshot
