Host-held state means the application, not the chat history, is the source of truth for what an agent has done. Several frameworks keep state on the host in some form. What decides whether the state stays right over many steps is five design questions. Below are those questions, how the AXI host-side runtime answers each one, and how to measure the same thing in any framework.
Naming note: "AXI" here is the AXI host-side runtime for LLM agents, an architecture by Rajnish Choubey. It is not the ARM AMBA AXI chip-bus protocol.
What is host-held state, and which frameworks have it?
Frameworks such as LangGraph (typed state with checkpointers), Letta (formerly MemGPT, host-managed memory), PydanticAI (typed dependencies and validated outputs) and smolagents (code agents whose variables persist in a host-side interpreter) all keep some agent state outside the message list. They are general tools, and each lets you build different things on top. AXI is one specific design with one purpose: keep an edited state exactly right while a model proposes changes to it, and it reports measured results for that. It can sit beside or under those frameworks.
The five questions
| # | Question to ask of any design | How AXI answers it | How to measure it in any framework |
|---|---|---|---|
| 1 | Who may change the state? | Only the host. The model proposes one change at a time in a short fixed format. | Count steps where the model's output touched state without passing a check. |
| 2 | What happens to a change that breaks a rule? | It is refused with a typed reason, and the state is left untouched. | Inject known-bad proposals, count how many reach the state. |
| 3 | What happens to a change that is almost right? | The host completes what the rules can complete (deleting a service also removes its edges) and then commits. | Take the model's real slips, count how many a fixed rule repairs. |
| 4 | Can a change half-apply? | No. A change commits all or nothing. | Kill the process mid-commit, audit the state. |
| 5 | What does the model see next? | The exact state, not the history. | Log the prompt size at step 1 and at step 10,000. |
What did those answers measure?
- Question 3, completion. Gemini 3.8 Flash, 4,000 requests: the model's replies were right 3,954 times (98.85 percent). All 46 slips were of one kind, a delete that left out the service's other connections. A fixed rule completed all 46. A gate that only refuses (question 2 without 3) kept the state well formed but exactly right on only 84 of 400 steps; with the completion it was 400 of 400.
- Question 1 and 2, the check. 10,000,000 proposals against the Python reference model on a CPU, 1,401,041 of them injection attempts: none of the injection proposals committed, and the held state matched an independent reference at every comparison. That run also found four things to harden (listed on the test page).
- Question 5, the prompt. The prompt stayed near 620 tokens and never above 660 over four jobs of 20,000 steps on a 7B model, with the state exactly right on 80,000 of 80,000 steps. A kept chat of the same job grew until it hit the window.
- Across models. The host's state was exactly right in the main cell of 10 of 11 open models, from 1.5B to 32B, with 0 corrupted steps in every cell. The eleventh, a DeepSeek-R1 distill, ran every reply to a delete to the 96-token cap; the host dropped those replies, so its state stayed well formed but fell behind (1.53 percent exact).
- Other kinds of state. A table, a schedule, source code and tool calls, on a 7B and a 14B: with the world's own rule check, the held state broke no rule on 3,600 steps.
What does this not claim?
The check can only enforce rules that are written down. It decides whether a change is legal, not whether it was the right change. The measured work uses synthetic service graphs and four small synthetic worlds.
Run it
git clone https://github.com/gaurshalendra2412/axi-runtime
cd axi-runtime/python
python3 quickstart_replay.py
No key, no GPU. It replays the saved replies of the 4,000-request run and prints 4,000 of 4,000, 3,954 of 4,000, 85 of 400 and 4.5 times. Test pages: github.com/gaurshalendra2412/axi-runtime. White paper: doi.org/10.5281/zenodo.23958737. Ten open-model tests: doi.org/10.5281/zenodo.24319561.
Architecture by Rajnish Choubey. Tests built and run by Shalendra Gaur.
Top comments (0)