Statelet is an Apache 2.0 agent runtime data layer for production AI systems, written in Rust. It unifies durable memory, runtime state, declarative memory queries, DiskHNSW/SPFresh vector recall, temporal graph traversal, CDC, Raft consensus, Redis compatibility, and a management UI in one engine.

How fast is Statelet's temporal graph engine?

Measured in July 2026 on Apple Silicon with 10K nodes and 50K temporal edges, Statelet ingests edges at 3.8 microseconds per edge, runs temporal range queries in 0.3 microseconds, depth-3 BFS in 13.8 microseconds, and point-in-time AS OF queries in 0.6 microseconds.

What memory accuracy does Statelet reach?

Statelet's server-side memory engine was evaluated on LongMemEval-S and reached 91.6% full-pipeline accuracy over 500 questions, with 99.0% hit@5 retrieval recall. Search uses local embeddings and vector plus graph retrieval; the LLM is only used during ingestion when enabled.

What is temporal graph recall in Statelet?

Statelet stores memories as graph nodes and temporal edges with validity windows. You can query what was true at a specific time with AS OF, expand semantic vector hits through graph relationships, and follow supersedes chains for updated or conflicting memories.

What languages and protocols does Statelet support?

Statelet provides gRPC SDKs for Python (3.9+), Rust (1.70+), Go (1.21+), Java (17+), and C++ (17+), plus an npm memory plugin for coding agents. It also natively supports the Redis RESP2 protocol, so existing Redis clients can connect directly on port 6379.

How does Statelet isolate enterprise tenants?

Statelet isolates tenants with a namespace to database hierarchy. The selected scope is baked into physical key, vector index, and graph names before shard routing, so agents or end users can own identical logical names without sharing a physical key space.

What vector search capabilities does Statelet have?

Statelet includes DiskHNSW and SPFresh vector search backends with L2, cosine, and inner-product distance metrics. It supports Product Quantization and Scalar Quantization for memory-efficient search, sparse hybrid retrieval, graph-RAG expansion, and distributed fan-out search across cluster nodes.

How is Statelet deployed?

Statelet can be run from source with three services: metadata_service, raft_engine with the data-node feature, and gateway. It also ships Kubernetes manifests and macOS launchd service definitions. The topology consists of stateless gateways, a metadata Raft plane, and sharded data nodes.

What interfaces does Statelet expose?

Statelet exposes the Statelet gRPC service, a Redis RESP2-compatible TCP server, REST management APIs on the gateway, a React management UI, and a durable SubscribeCommitted CDC feed for catch-up and live-tail integrations.

Is Statelet open source?

Yes, Statelet is fully open source under the Apache 2.0 license. The source code is available on GitHub at github.com/stateletlab/statelet.

Statelet - Agent Runtime Data Layer

Name: Statelet
Availability: InStock
Author: stateletlab

Statelet - Agent Runtime Data LayerStatelet is an Apache 2.0 distributed database written in Rust. It gives AI agents durable memory, context versions, runtime state snapshots, an openCypher-style memory query language, DiskHNSW/SPFresh vector recall, temporal graph recall, CDC, Raft consensus, Redis compatibility, and local ONNX embeddings.Statelet Features

Unified Runtime Data Engine - Memory, context, state snapshots, vector recall, temporal graph, and triple-store rows share one Rust engine and transaction model.

Memory Query Language - Query agent memory declaratively with openCypher-style syntax, including vector search, graph expansion, and AS OF time travel.

LSM-Tree Storage Engine - Concurrent skip-list MemTable, prefix-compressed SST files, bloom filters, sharded LRU block cache, leveled compaction.

DiskHNSW / SPFresh Vector Search - Dense, sparse hybrid, graph-RAG, and quantized vector retrieval with local ONNX embeddings.

Temporal Graph Recall - Time-aware edges, AS OF queries, supersedes chains, and graph expansion for explainable memory.

Multi-Tenancy - Namespace and database scoping physically isolates keys, indexes, and graphs for every tenant or agent.

CDC Feed - Durable SubscribeCommitted stream keyed on Raft log offsets for catch-up and live-tail integrations.

Raft Distributed Consensus - Separate metadata and per-shard data Raft groups for production topology.

Redis Protocol Compatible - RESP2-compatible TCP server. Any Redis client can connect directly.

Multi-Language SDKs - Official gRPC clients for Python 3.9+, Rust 1.70+, Go 1.21+, Java 17+, and C++ 17+.

Management Web Console - Namespaces, databases, graph queries, KV operations, user management, and Prometheus metrics.

ArchitectureThree-tier distributed architecture: stateless Gateway (gRPC :9379, HTTP :9380, Redis :6379) routing to a Metadata Plane (Raft consensus for namespace/database registry, shard map, column families, and node registry) and Data Plane (ShardEngine per node with LSM-tree, DiskHNSW/SPFresh vector search, temporal GraphSST, and per-shard Raft).PerformanceTemporal graph (measured 2026-07, Apple Silicon, 10K nodes / 50K edges)

Edge ingestion: 3.8 microseconds per edge, about 264K edges per second

Temporal range query: 0.3 microseconds

Multi-hop BFS depth 3: 13.8 microseconds

Point-in-time AS OF query: 0.6 microseconds

LongMemEval-S memory evaluation

Full-pipeline accuracy: 91.6% over 500 questions

Retrieval-only hit@5: 99.0%; recall@5: 0.937; MRR: 0.938

span_hit@5: 0.900; span_recall@5: 0.789; top-1 similarity: 0.994

End-to-end overall: 458 / 500 correct; strongest categories are single-session-assistant at 98.2% and single-session-user at 97.1%

Search is pure vector plus graph retrieval; LLM extraction is ingestion-time only when enabled

Open source and Apache 2.0 licensed.Agent MemoryStatelet's gateway extracts structured facts from conversations, detects conflicts through vector similarity, resolves updates with temporal supersedes edges, and stores each fact with temporal metadata. Agents can recall memories with vector search, hybrid dense+sparse search, graph-RAG expansion, Personalized PageRank, and AS OF time travel.SDKs & ProtocolsOfficial gRPC clients: Python 3.9+, Rust 1.70+, Go 1.21+, Java 17+, C++ 17+. Redis RESP2 protocol compatible — connect with redis-cli or any Redis client library.DeploymentSingle Node (cargo run), Docker (multi-stage build), Kubernetes (StatefulSet with 3+ data nodes), macOS launchd.Open source, Apache 2.0 licensed. View on GitHub.