Superbrain is a next-generation distributed memory fabric designed for AI agents. It transforms passive storage into a cognitive participant in your agentic workflows.
Traditional databases treat data as static bytes to be retrieved. Superbrain treats data as Active Memory—with semantic awareness, intentional decay, and microsecond-latency sharing between co-located and distributed agents.
Superbrain separates the **Control Plane** (Metadata, Consensus, Routing) from the **Data Plane** (Raw Bytes, mmap, rDMA). This allows for extreme horizontal scale without impacting metadata consistency.
Hot Path Bypass: Once the SDK resolves metadata from the Coordinator, subsequent Reads/Writes go directly to Fabric Nodes.
Active Memory transforms distributed storage into a cognitive participant. Unlike Redis or S3, Superbrain understands the "liveliness" and "intent" of the data it stores.
When multiple agents write to the same context, Superbrain resolves logical conflicts using "Intent-Based Locking" rather than just row-level locks.
Set a decay_rate on any write. Low-priority reasoning steps will naturally "evaporate" from RAM, while high-priority terminal decisions persist.
Get your agent swarm up and running in minutes.
pip install superbrain-sdknpm install superbrain-distributed-sdkgo get github.com/golightstep/superbrainSdkfrom superbrain import SuperbrainClient
client = SuperbrainClient("localhost:50050")
# Write to distributed RAM
ptr_id = client.allocate(1024)
client.write(ptr_id, 0, b"Agent State Alpha")const { SuperbrainClient } = require('superbrain-distributed-sdk');
const client = new SuperbrainClient('localhost:50050');
const ptrId = await client.allocate(1024);
await client.write(ptrId, 0, Buffer.from("Agent State Alpha"));client := sdk.NewClient("localhost:50050")
ptr, _ := client.Allocate(1024)
client.Write(ptr.ID, 0, []byte("Agent State Alpha"))Tune every aspect of the SDK and Cluster behavior. Superbrain supports configuration via Environment Variables (recommended for containers) or explicit `Config` objects.
| Parameter | Default | Description |
|---|---|---|
SB_COORDINATOR | localhost:50050 | Primary gRPC endpoint. Comma-separate for HA pools. |
SB_ENCRYPTION_KEY | null | 32-byte key for E2EE. Enabling this wraps all IO in AES-GCM. |
SB_BYPASS_LOCAL | true | If node IP is 127.0.0.1, use SHM for 15μs latency bypass. |
SB_MTLS_ENABLE | false | Force mutual TLS certification for all agent operations. |
Superbrain provides three primary API tiers. Level 1 is the Raw Pointer API, offering maximum performance and control over byte offsets.
Reserves chunks of RAM across the distributed fabric. Returns a 36-byte UUID pointer.
# Allocate 50MB
ptr_id = client.allocate(50 * 1024 * 1024)
print(f"Memory ready: {ptr_id}")// Allocate 50MB
const ptrId = await client.allocate(50 * 1024 * 1024);
console.log(`Memory ready: ${ptrId}`);// Allocate 50MB
ptr, err := client.Allocate(50 * 1024 * 1024)
fmt.Printf("Memory ready: %s\n", ptr.ID)size | uint64 | Number of bytes to allocate. |
ERR_QUOTA_EXHAUSTED: Cluster has no free RAM in this tier.ERR_COORD_UNREACHABLE: Metadata resolve failed.Enriches raw data writes with semantic metadata. This is the foundation of Superbrain's "Cognitive Architecture".
summary | string | LLM-readable summary of the chunk content. |
intent | string | The logical purpose of this write (e.g. "Drafting Plan"). |
liveliness | float (0-1) | Sets the decay rate. 0.1 = ephemeral, 1.0 = persistent. |
One agent downloads a 500MB PDF. Instead of sending the PDF to 10 other agents, it writes to Superbrain and shares the 36-byte pointer via a message bus.
An execution agent is writing logs. A security agent subscribes to tag="Sensitive". The moment the executor writes an API key by mistake, the security agent is woken up by Superbrain to scrub the memory and flag the breach.
A Go ingestion agent writes high-speed telemetry. A Python agent reads the same offsets for ML inference. Superbrain handles the byte-order and mmap alignment automatically.