Persisting

Persistent Infrastructure for the Agent Era.

From model state to Agent history.

Persisting connects durable model state—parameters and KV caches—with durable Agent history—trajectories and execution records. The positioning is broader than any single command; the current product provides two concrete workflows:

• pvisor runs one Agent in a controlled environment and lets you review its effects before accepting them;
• pchronicle browses, queries, exchanges, and serves trajectory Datasets.

Use either workflow on its own. pChronicle can read external trajectory data without pVisor, and pVisor does not need pChronicle to complete its local run-review-apply loop. When connected, they preserve a path from execution to queryable history.

Install

The Python wheel installs the Python package and its matched public command-line entry points:

pip install persisting

pvisor --version
pchronicle --version

The rolling nightly build is available without a Rust toolchain:

curl -fsSL https://raw.githubusercontent.com/DeepLink-org/Persisting/main/scripts/install-nightly.sh | bash

Source developers can install the same command set with just install-cli. See the installation guide for platform requirements and executor setup.

Choose a workflow

Run one Agent and review its changes

pvisor run --safe codex
pvisor review last
pvisor apply last --all # or: pvisor drop last

--safe stages workspace changes. The exact filesystem and network boundary is platform-dependent and is recorded with the Run; consult the pVisor guide before treating it as a security boundary. This workflow does not require pChronicle.

Query trajectory history

pchronicle onboard
pchronicle onboard query
pchronicle agent codex ./trajectory-data \
  --ask "Which tools fail most often?"

The onboarding flow creates a temporary example Dataset and does not require a source checkout. A Dataset may be a local path, an object-store URI prefix, or a user alias such as @prod. pchronicle import accepts ATIF, ACTF, and OpenAI Messages; pchronicle export supports those formats plus Storyline JSON. pchronicle agent launches Codex or Claude with an ephemeral Dataset analysis skill. It instructs the Agent to use read-only pChronicle commands without changing the Agent's existing filesystem, network, or tool permissions. pchronicle serve starts the loopback-only, read-only Dataset UI and API.

pChronicle performance

Criterion.rs microbenchmarks and hyperfine lifecycle scenarios are compared against main in CI. Measurements use stable JSONPath addresses in the raw JSON report. See the benchmark runner and report contract.

No nightly benchmark has been published with the unified report format yet.

Command ownership

Command	Primary responsibility
pvisor	One Run, environments, review, checkpoints, apply/drop
pchronicle	Dataset catalog, SQL, built-in analysis, find, import/export, read-only serving

Current maturity

Capability	Status
pVisor host execution, review, checkpoints, and transactional workspace	Implemented
pChronicle local/S3 catalog, bounded SQL, analysis, find, import/export	Implemented
pChronicle loopback-only read API and embedded Web UI	Implemented
Gateway capture and cooperative proxy policy	Implemented
Container/libkrun executors and transparent network boundaries	Platform-dependent; see the pVisor and OverlayNet docs
Queue and document Search	Separate stable capabilities
Tensor Memory / TTAS	Experimental

The CLI --help, component READMEs, tests, and user guides describe supported behavior. Files under docs/src/design/ may also contain explicitly labelled target architecture; RFCs preserve decisions and are not command references.

Documentation

• Choose a workflow
• Run your first Agent
• pVisor documentation
• Review and selectively apply changes
• pChronicle documentation
• Explore durable history
• Project architecture
• Runnable examples

From a source checkout:

just examples
just docs-build
just docs-links

License

Apache License 2.0. See NOTICE for third-party attributions and separately licensed bundled components.