Persisting

Persistent infrastructure for the Agent era.

Persisting connects durable model state—parameters and KV caches—with durable Agent history—trajectories and execution records. The current product is the path from execution to queryable history:

• pvisor is an executor that produces persistable, reviewable facts: staged Effects and execution records from one Agent Run;
• pchronicle browses, queries, exchanges, and serves trajectory Datasets.

Each command works on its own. Connected, they cover pvisor run --safe → review/apply → configured capture → a queryable Dataset.

Install

pip install persisting[lance]
pvisor --version
pchronicle --version

The rolling nightly build installs the same commands without a Rust toolchain:

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

See the installation guide for platform requirements and executor setup.

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; nothing enters your project tree before you accept it. The exact boundary is platform-dependent and recorded with the Run—consult the execution guide before treating it as a security boundary.

Query Agent 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—no source checkout required. pchronicle import accepts ATIF, ACTF, and OpenAI Messages; pchronicle serve starts a loopback-only, read-only Dataset UI and API.

After capture is configured, selected pVisor Run events can enter a pChronicle Dataset. See the capture guide.

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

Documentation

• Choose a workflow — pick the entry point that matches your task
• Run your first Agent — the run-review-apply loop
• Explore durable history — browse and query a trajectory Dataset
• Project architecture — ownership and delivery boundaries

Criterion.rs microbenchmarks and hyperfine lifecycle scenarios are compared against main in CI; see the benchmark contract.

Latest nightly pChronicle benchmark: 073a70bdf4ff on linux/x86_64 (2026-09-02T15:21:52.794687+00:00).

Case	Metric	Value
criterion/atif_conversion/parse_corpus	latency_median_ns	4.979e+06 ns
criterion/atif_conversion/roundtrip_corpus	latency_median_ns	6.552e+06 ns
criterion/projection_cpu/events_to_storyline_corpus	latency_median_ns	2.968e+05 ns
system/projection_pipeline/event_append	initial_append_ms	66.059 ms
system/projection_pipeline/projection_build	build_ms	5765.31 ms
system/projection_pipeline/projection_incremental	sync_ms	44.614 ms
system/lance_vs_json/lifecycle	cold_query_ms	3666.153 ms
system/lance_vs_json/lifecycle	get_storyline_full_ms	10.341 ms
system/lance_vs_json/lifecycle	replace_storyline_ms	44.332 ms
system/lance_vs_json/selective	lance_qps	310.7 ops/s
system/lance_vs_json/group_by	lance_qps	423.4 ops/s
system/lance_vs_json/summary	lance_over_json	0.244 ratio
system/json_streaming_ndjson/json_streaming	p95_ms	13.891 ms
system/json_streaming_ndjson/json_streaming	rows_s	3.055e+05 ops/s
system/json_streaming_ndjson/json_streaming	process_peak_rss_mib	36.508 MiB
hyperfine/projection_pipeline	wall_median_seconds	5.961 s
hyperfine/lance_vs_json	wall_median_seconds	47.511 s

Open the complete benchmark run.

License

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