Hybrid search, graph indexing, and vector-symbolic memory for AI agents
Agent harness infrastructure that gives AI agents memory for each user, connected accounts, action limits and an audit record.
Stage
Startup
Not given
Founded
2026
Not given
Based in
US
CA
Pricing model
Open source
Freemium
Pricing
Open source under Apache-2.0 license. WARM Cloud is in development with usage-based pricing for sync, backup and verification; details to be published at launch.
Free: first 500 monthly-active end users per month, free forever. Paid: $0.30 per monthly-active end user above 500, billed monthly.
Key features
Hybrid search combining BM25 lexical and vector scoring
Vertex graph vector index with 2.29-2.75x throughput vs hnswlib
Vector-symbolic fact memory with holographic reduced representations
AES-256-GCM encryption at rest with Argon2id key derivation
Signed audit log with Ed25519 and W3C DID identity
Secure purge that overwrites deleted data with zeros
Crash-safe append-only log with checksumming
Native Rust API with Python HTTP client and Node.js bindings
MCP server integration for editors and AI agents
WARM Cloud sync, backup and verification without server access to keys or documents
Timestamped memory ingestion with backdating capability
Per-user data isolation and privacy scoping
Ranked memory recall with dated evidence
Connected accounts integration per user
Long-running agent jobs with folder-level operations
Configurable token ceiling per API call
Enforced action limits on agent operations
Complete action audit trail with hash chaining
MCP server integration for Claude and other clients
What makes it different
Runs entirely locally in your process with no network calls or telemetry
Encryption at rest built into the open-source engine itself, not an enterprise add-on
Significantly faster vector search than hnswlib with graph-indexed vertex retrieval
Signed provenance and verified audit trail using C2PA, COSE and SCITT standards
Embedded library model eliminates dependency on external services or vendor lock-in
Highest published BEAM benchmark score at 128K tokens: 85.1, compared to competing memory systems
98% fewer tokens required than raw tool output, making responses 4× cheaper
Works with 20+ agent frameworks including LangChain, LangGraph and LlamaIndex
First 500 monthly-active end users free forever
Per-person billing model scales with user growth, not API calls
warm-rust
Harny