ciris-persist 9.4.0

Unified Rust persistence for the CIRIS federation — signed events, time-series, runtime state.

CIRISPersist

One embeddable Rust crate behind every stateful surface a CIRIS federation node needs — the signed reasoning-trace log, the hash-chained audit log, the memory graph, time-series telemetry, secrets-at-rest, federation trust state, and the CEG §19 holonomic / forever-memory substrate (fountain content, WholenessWitness, aggregation tiers). Postgres or SQLite; in-process via PyO3 or over HTTP.

Current: v8.5.0 · CIRISVerify pin v5.10.0 · CEG §19 / §19.7 (1.0).

What it is

CIRISPersist is the lowest stateful substrate above CIRISVerify. An agent or lens node links it as a library — pip install ciris-persist — and gets every storage surface it needs from one versioned API instead of ~11 hand-rolled services. The backend is chosen at Engine construction by DSN scheme (postgres://… or sqlite://…); every method works on both.

Substrates

Substrate	Backs
trace ingest	signed reasoning-trace event log + LLM-call log
cirisaudit	hash-chained, per-tenant signed audit log (RFC 6962 Merkle)
cirisgraph	memory nodes + edges — absorbs MemoryService / ConfigService
telemetry	metric writes + TSDB rollup
secrets	federated SecretsService — AES-256-GCM at rest, hardware-backed master key
cirisnode	CIRISNodeCore federation-consensus substrate
fountain	fountain-coded content primitive — signed manifest + N+K opaque symbols, tier eviction, authenticated partial reads (CEG §19.3)
wholeness_witness	§19.1 WholenessWitness corpus — divergence detector that routes to the existing quorum-merge (CEG §19)
content_aggregation	§19.7 forever-memory aggregation tier — N→1 composites, opaque aggregation_meta, descent-integrity (CEG §19.7)
sequence / occurrence	atomic per-identity counters + endpoint-liveness registry
lens substrates	tasks, thoughts, correlations, tickets, deferral reports, WA certs, …

Honest read

• Postgres + SQLite at 100% parity. Every PyO3 method works on both backends — including the observability read API and the lens-derived schemas — so sovereign-mode (Pi / iOS) deployments are not second-class.
• In-process cohabitation. Engine is a process-singleton: a CIRIS 3.0 process hosting the agent + NodeCore + LensCore shares one runtime, one pool, one identity — see docs/COHABITATION.md. The public ffi::pyo3::register(py, m) hook (#231) lets a host wheel re-export the whole persist surface into its own module — one .so, one PyO3 type registry (CIRISServer one-wheel).
• Hardware-backed secrets. The secrets master key is derived — via CIRISVerify — from a seed sealed by the platform TPM / Keystore / Secure Enclave where one exists, with an honest software fallback where it does not.
• Crypto goes through CIRISVerify. Persist never rolls its own — signing, verification, and key derivation route through ciris-verify-core / ciris-keyring (pinned v6.0.0, all six crates flipped in lockstep). The §19 holonomic gates (verify_witness, compare_witnesses, verify_aggregation_meta, verify_member_commitment, ejection_verdict, compute_merkle_root) are all verify-core calls — persist never re-rolls Merkle / preimage / signature logic.
• Holonomic / forever-memory substrate (CEG §19 / §19.7, 1.0). Persist is the store + WholenessWitness-corpus owner + the divergence→quorum-merge router. A withdrawn content_id is hard-deleted regardless of rarity (revocation overrides rarity); a witness equivocation is retained and flagged hard_case:witness_equivocation, never reconciled. Memory fades along one pressure-driven descent but cannot be falsified — the signed envelope/manifest is the incorruptible anchor and partials stay authenticated against the signed per-symbol hashes.
• Constitution alignment (0.1.5). v9.0.0 binds persist to the substrate clauses of the CIRIS Constitution 0.1.5. Federation-tier ingest is PQC-mandatory — every tier = federation write is hybrid-verified (Ed25519 + ML-DSA-65, Strict) at the admission gate, local-tier exempt (CC 5.3.2.4.3.1; BREAKING — classical-only federation rows are non-conformant). A node-only key's delegates_to may carry only infra:* scopes, never agency:* — "infrastructure must not have agency" enforced cryptographically (CC 4.4.3.4.3 / CC 1.13.5). Non-infra community membership for a node/agent key requires owner-binding to a human (CC 3.2 / CC 3.4.7.1), and community content is sealed under a per-community DEK rotated zero-window on member removal (CC 4.4.3.2.1 / 4.4.3.2.2, wrap_algorithm: v2 only). Privacy stays bounded — cohort scope hides content, not contact (CC 1.13.3.1; see MISSION.md §1.9). See docs/THREAT_MODEL.md §3.15 (AV-67..AV-72).
• 1+4 lockdown. Persist adds zero structural primitives (CC 1.7 — the federation's one workhorse attestation primitive + four structural composers). Every substrate above rides scores + a subject_kind + the four composers (delegates_to / supersedes / withdraws / recants) — the constitution-alignment gates are admission logic over the existing federation_keys / attestation surface, not new wire shapes.
• Deliberately not: no embedded graph DB engine (Postgres / SQLite recursive CTEs instead); not a daemon (a library, not a service); horizontal sharding is out of scope.

Performance & SOTA

Measured by the in-repo criterion suite (benches/ — eleven harnesses run per commit and published to the trend dashboard). Representative current numbers:

• AES-256-GCM (secrets-at-rest) — ~9.5 GiB/s, via CIRISVerify v2.8.0.
• Analytics — V042 covering indexes turn the scoring ReadEngine queries into index-only scans; cross_agent_divergence ~−42% vs. a raw table scan.
• next_sequence — a durable, async-safe atomic counter increment ~10 µs (the SQLite UPSERT itself ~2 µs; the rest is the async wrapper).
• Cold start — open + full migration run ~12 ms.

Capability	SOTA peers	CIRISPersist
Embedded persistence	ORMs — sqlx, Diesel; per-service DBs	one API, Postgres + SQLite at 100% parity — at parity
Audit log	Trillian, Rekor, AWS QLDB	RFC 6962 Merkle + post-quantum-signed tree heads — ahead of typical
Crypto-at-rest	ring, OpenSSL (~3–6 GiB/s AES-GCM)	~9.5 GiB/s AES-256-GCM via CIRISVerify — ahead
Post-quantum	mostly classical	hybrid Ed25519 + ML-DSA-65 throughout; Full-mode traces + fountain manifests + aggregation meta reject classical-only at ingest (HNDL hard cut, CEG §10.1.5.1.1) — ahead
Analytics	DuckDB, ClickHouse (columnar)	row store + query-shaped covering indexes ("poor-man's column store") — at parity for the fixed query set; ~2–5× behind a columnar engine on raw ad-hoc scan
Horizontal scale	sharded DB services	a library, not a service — behind (deployment shape, not algorithm)

Ahead on post-quantum crypto and the Merkle audit log; at parity on embedded persistence and on analytics for its closed query set; the one real "behind" — horizontal scale — is a library-vs-service choice, not an algorithm gap.

Quick start

import ciris_persist as cp

engine = cp.Engine(dsn="sqlite://./agent.db", signing_key_id="agent-ed25519")
engine.register_consumer("my-adapter", ["cirisgraph"])
summary = engine.receive_and_persist(request_body_bytes)

Docs

Doc	What
MISSION.md	Mission-Driven Development alignment (Accord Meta-Goal M-1)
FSD/CIRIS_PERSIST.md	Full functional spec
docs/COHABITATION.md	In-process cohabitation model
docs/THREAT_MODEL.md	Threat model (AV-* attack vectors)
docs/PUBLIC_SCHEMA_CONTRACT.md	Stable schema contract
CHANGELOG.md	Per-release history

License

AGPL-3.0-or-later. The persistence path is auditable line-by-line by design: closed-source forks are forbidden, which makes the federation primitive's audit story structurally enforceable, not merely socially expected.