Current reference semantics
Storage protocols and conformance
Benchmark first
Corpora: BEIR, FEVER, and LongMemEval form the end-to-end matrix; deterministic fixture suites cover stores, serialization, compiler output, and failure recovery.
Protocol: Pin every input, dependency, model, prompt, and seed. Preserve per-case traces and report quality beside p50/p95 latency, tokens, storage, index-build time, refresh work, and failure count. Compare outputs after restart and across supported stores; compiled configurations must reproduce the same artifact graph and evaluator inputs.
InMemoryArtifactStore is the executable reference for optimistic revision checks, tenant isolation, explicit supersession, history, and point-in-time reads. Production adapters can run the same behavioral cases.
How it works
Protocols specify observable behavior rather than backend classes. A store implementation declares capabilities, then runs a shared conformance suite against replay, atomic revision, isolation, deletion, deterministic ordering, and point-in-time cases. Cross-store operations use an application transaction/outbox boundary; Mari does not pretend separate databases share an atomic commit. Indexes remain disposable projections that can be rebuilt from documents and artifacts.
Research basisInvariant confluence shows that safe coordination depends on application invariants. Mari therefore specifies atomicity, replay, isolation, time-travel, and deletion behavior independently of backend methods. The protocol split is library design, not a result asserted by the paper.
from mari_components.platform import InMemoryArtifactStore, RevisionConflict
store = InMemoryArtifactStore()
store.commit(first_revision, expected_revision=None)
try:
store.commit(second_revision, expected_revision="stale-revision")
except RevisionConflict:
refresh_and_retry()
current = store.get("fact:refund-window", scope=scope)
historical = store.at_time(
"fact:refund-window",
scope=scope,
known_at=query_time,
)
The bundled conformance tests cover replay safety, deterministic ordering, point-in-time reads, tenant isolation, and atomic revision checks. Physical deletion and cross-database transactions remain adapter-specific capabilities and must be tested by the host backend.