Real runtime
MLX to vLLM
Local Apple Silicon source, AWS g4dn.xlarge target, vLLM 0.23.0, Qwen2.5 0.5B, complex graph/KV binding aligned, then graph activity resumed on the target.
Agents can move: validated round trip
Move the mind,
not just the model.
PermeantOS is an open-source platform for live AI agent state migration. Validated runs show a graph-attached agent can move from Apple Silicon MLX to AWS vLLM with live KV cache, artifacts, memory, retrieval evidence, and pending tool policy intact; resume activity on the target; export the target decode boundary back through a live vLLM API; import it into MLX; and continue again at the origin.
Current validated result
Local Apple Silicon MLX to AWS g4dn.xlarge vLLM 0.23.0, using Qwen/Qwen2.5-0.5B-Instruct. The strongest run validated a 128-token continuation horizon, a 27-node Agent Memory Graph package, target-side graph activity, reverse vLLM-to-MLX runtime import, and origin return-home continuation. A local MLX-to-llama.cpp proof now shows canonical KV tensors from one runtime can feed directly into another runtime's raw KV writer.
- 2016
- long-horizon prefix tokens
- 24
- Qwen layers migrated
- 27
- graph nodes moved
Permeant migration node
USXF v1.1 SOURCE
Apple Silicon MLX
live source adapter
signed USXF stream
TARGET
AWS vLLM CUDA
prefix-cache attachment
Interactive migration trace
Live Migration Simulator
Step through the same handoff PermeantOS performs in the validated MLX-to-vLLM path: freeze, encode, stream, seed, verify, commit.
SOURCE HOST
Apple Silicon / MLX
% Encoded
USXF sealed frame
TARGET HOST
AWS g4dn / vLLM
PERMEANT DAEMON TELEMETRYMLX-TO-VLLM-X64
Telemetry ready. Click "Migrate State" to capture live attention caches.
USXF v1.1
A runtime-neutral envelope for active KV state.
USXF stores model identity, attention geometry, sequence length, dtype, transfer quantization, block hashes, checksums, and signatures so source and target runtimes can reason about migratable state without sharing a hardware layout.
The current implementation validates live KV-cache migration plus Agent Memory Graph transaction binding, target-side activity resume, reverse vLLM-to-MLX runtime-state import, and origin return-home graph continuation for the MLX-to-vLLM path.
{
"usxf_version": "1.1",
"source_runtime": "mlx",
"target_runtime": "vllm-0.23.0",
"model_architecture": "Qwen2.5-0.5B-Instruct",
"sequence_length": 2016,
"layers": 24,
"block_hashes": ["sha256:..."],
"signature": "ed25519:..."
}A compact USXF envelope view: identity, runtime geometry, hashes, and signature in one migratable contract.
Evidence, not theater
Validated migration evidence and platform path.
The strongest result is precise and reproducible: real complex-agent MLX-to-vLLM migration, exact 128-token continuation fidelity on Qwen2.5, target-side Agent Memory Graph continuation, reverse runtime import back into MLX, origin return-home proof, and local MLX-to-llama.cpp canonical KV feed evidence.
Validation
Exact 128-token horizon
Hash validation passed, post-migration decode matched source and target-baseline continuation through the 128-token Qwen2.5 horizon, the target exported its decode boundary, MLX continued from that target-advanced state, and the origin continued from the AWS graph proof.
Analysis
Codec roadmap
Historical QATQ runs showed promising transfer reduction, but QATQ is being matured separately. PermeantOS now treats raw and FP8 as core paths, with QATQ folded back in when it is ready as a proper crate/service.
Latency Savings Calculator
Uses the whitepaper prefill model:
T = alpha*S + beta*S^2, with alpha=1e-5 and beta=6e-10. This is analytical, not a measured E2E runtime.
Analytical re-prefill estimate0.97s
Validated live migration result2016-token prefix
Analytical comparison
Long contexts are where state migration starts to matter. Actual cutover time depends on model geometry, bandwidth, quantization, runtime hooks, and host warmup.
Architecture
How PermeantOS works.
The daemon and adapters coordinate extraction, normalization, transport, target allocation, verification, and commit.
Stage 01
Roadmap
Toward a repeatable open-source platform.
KV cache portability is the first layer, and the agent movement claim is now validated for specific runtime paths. The next milestone is product maturity: stable releases, binaries, crates, documentation hub, automated evidence jobs, durable graph sessions, and broader runtime coverage.
KV cache migration
Validated MLX source extraction, USXF transport, Agent Memory Graph binding, target KV write, prefix-cache seeding, exact short-horizon continuation, reverse runtime import, AWS target-side graph activity resume, and origin return-home proof.
Validated
Repeatable platform releases
Add GitHub Releases, signed binaries, checksums, crate publishing readiness, SDK packaging, docs hub, and automated evidence jobs.
Next
Broader runtime fidelity
Repeat validation across longer horizons, larger contexts, more model families, more runtime adapters, and cost-bounded cloud evidence jobs.
In progress
Apache 2.0
Contribute to portable long-running intelligence.
PermeantOS is open-source infrastructure for portable agent state. Contributions are especially welcome around runtime adapters, manifest analysis, Agent Memory Graph export/import, reproducible benchmarks, security review, release packaging, and documentation.
Technical whitepaperJune 2026
PermeantOS White Paper: Live migration for AI agent state
A technical walk-through of USXF, graph-attached MLX-to-vLLM validation, target-side continuation, reverse runtime import, origin return-home proof, MLX-to-llama.cpp evidence, and the roadmap toward a repeatable open-source platform.
HTML documentOpen