superbrain-sdk 0.7.5

High-performance distributed memory fabric for AI agents.

🧠 superbrain-sdk v0.7.5 — Python

SuperBrain is a high-performance distributed memory fabric that allows multiple AI agents to share state and context at microsecond latencies.

🚀 Viral Integration (CrewAI)

The easiest way to share context across your CrewAI swarm. Just decorate your task creation:

from superbrain import shared_context
from crewai import Agent, Task

@shared_context("market-research")
def create_task(ctx, researcher):
    # This task's context is now live on the SuperBrain fabric
    return Task(description="Analyze 2024 AI trends", agent=researcher)

Now with Phase 3: Automated AI Memory Controller, LangChain & PyTorch integration, and self-healing KV cache pooling.

---

🚀 What Is SuperBrain?

SuperBrain is a distributed RAM network where multiple AI agents on different machines share memory via 36-byte UUID pointers — instead of copying massive JSON blobs over slow APIs.

Key Numbers:

• ~100 MB/s write throughput per node (gigabit saturation)
• ~1–2ms read/write latency on LAN
• 36 bytes to share any amount of memory between agents
• Zero-copy context passing for multi-agent workflows

---

📦 Installation

pip install superbrain-sdk

---

✨ New in v0.7.1 — Tiered Architecture & Zero-Copy SHM

SuperBrain now operates as an ultra-fast L1 Shared Memory Tier for agent architectures.

• Coordinator Bypass: Metadata is cached locally, eliminating the gRPC hop to the Coordinator for established pointers.
• Zero-Copy SHM: When the SDK detects a co-located Memory Node (127.0.0.1), it seamlessly switches from gRPC streaming to direct /dev/shm memory-mapped file access.
• 13.5µs Native Latency: The Native Go core bypass achieves microsecond speed, while the Python SDK currently hits ~9ms due to CGo/ctypes FFI overhead.

# L1 Shared Memory via Circular Buffer
from superbrain.kv_pool import CircularBuffer

# Pre-allocated allocation-free Ring Buffer for ultra-fast Market Data ingestion
l1_stream = CircularBuffer(fabric, size=1024 * 1024)
l1_stream.push(b"AAPL 150.00") # Uses direct memory-mapped I/O if local

---

✨ Distributed Semantic Memory (v0.3.1 Features)

SuperBrain includes a production-ready, FAISS-backed Semantic MemoryStore that acts as a zero-network vector database. Instead of querying a remote database, SuperBrain pulls the entire FAISS index directly into your agent's local RAM instantly via the distributed fabric.

• 59μs Local Search: Once loaded, vector searches bypass the network entirely.
• Microsecond Memory Inheritance: Agents can 'inherit' the exact state of another agent's memory in ~6ms.

from superbrain.integrations.semantic import SemanticMemoryStore

store = SemanticMemoryStore(fabric, namespace="global-knowledge")
store.add("The capital of France is Paris", embedding=[...])

# Serialize FAISS index to distributed RAM
root_ptr = store.commit() 

# ---------------------------------------------------------
# ANY other machine can instantly clone this knowledge base:
# ---------------------------------------------------------
agent_b_store = SemanticMemoryStore(fabric)
agent_b_store.load(root_ptr) # <--- Inherited everything in ~6ms

# Network-free local search
results = agent_b_store.search(query_emb) # <--- Runs in ~59μs!

---

✨ Phase 3: Automated AI Memory Controller (v0.2.0 Features)

Zero-Config Cluster Discovery

from superbrain import AutoMemoryController

# Finds your SuperBrain cluster automatically via mDNS
memory = AutoMemoryController()

Shared Context Across Multiple LLMs

@memory.shared_context("research-session")
def researcher(ctx, document):
    ctx.write("findings", {"summary": "...", "confidence": 0.95})

@memory.shared_context("research-session")  # Same context!
def strategist(ctx, findings_ptr):
    return ctx.read("findings")             # Microsecond access

# Different LLMs, same shared memory:
researcher("War and Peace, all 1200 pages")
result = strategist(None)   # Claude reads what GPT-4 wrote!

Automatic KV Cache Deduplication

from superbrain import DistributedContextFabric

fabric = DistributedContextFabric(coordinator="localhost:50050")

# Same system prompts across 1000 agents → stored ONCE
ptr = fabric.store_kv_cache(b"You are a helpful assistant", model="gpt-4")
# Claude, Llama, and GPT-4 all reuse the same pointer

LangChain Memory Adapter

from superbrain.integrations.langchain import SuperBrainMemory
from langchain.chains import ConversationChain

memory = AutoMemoryController()
sb_memory = SuperBrainMemory(memory, session_id="user-123")

chain = ConversationChain(llm=your_llm, memory=sb_memory)
# Conversation history persisted in distributed RAM!
# Survives LLM restarts. Shared across machines.

PyTorch / HuggingFace KV-Cache Offloading

from superbrain.integrations.pytorch import enable_distributed_kv_cache

enable_distributed_kv_cache(fabric, max_local_layers=4)

# NOW: When GPU VRAM is full, KV caches page to cluster RAM
# instead of crashing or swapping to slow disk
model.generate(input_ids, max_length=100_000)  # Long context just works!

---

🔧 Core API

from superbrain import DistributedContextFabric
from superbrain.monitor import MonitorServer

# Initialize with all Phase 3 subsystems
fabric = DistributedContextFabric(coordinator="your-host:50050")

# Start live monitoring dashboard at http://localhost:9090
MonitorServer(fabric).start()

# Allocate + write data to distributed RAM
ptr = fabric.allocate_and_write(b"My huge AI context", agent_id="agent-1")

# Any machine anywhere can read it with just the pointer
data = fabric.read(ptr, 0, 0)

# Named shared contexts
ctx = fabric.create_context("agent-swarm")
ctx.write("state", {"step": 42, "done": False})
state = ctx.read("state")

# Get full telemetry
fabric.print_stats()

---

📊 Performance Telemetry

stats = fabric.stats()

# {
#   "telemetry": {
#     "throughput": {"write_mbps": 98.4, "read_mbps": 102.1},
#     "kv_cache": {"hit_ratio": 0.87},
#     "operations": {"write": {"p50_ms": 1.2, "p95_ms": 3.1, "p99_ms": 5.4}}
#   },
#   "kv_pool": {"total_segments": 142, "compressed_segments": 32},
#   "anomalies": []
# }

---

🔐 Zero-Trust Security

from superbrain.security import KeyManager, AnomalyDetector

# Per-context AES-256 key derivation
km = KeyManager(master_secret=os.urandom(32))
key = km.key_for("session-user-abc")
km.schedule_rotation("session-user-abc", interval_s=3600)

# Anomaly detection on access patterns (Z-score, 3σ)
det = AnomalyDetector()
# Automatically alerts when an agent accesses 100x more bytes than normal

---

🧹 Memory Management — When to Call free()

TL;DR — Use SharedContext or store_kv_cache() and you never need to call free().

What you call	Need free()?	Best for
client.allocate()	✅ Yes	Raw low-level control
ctx.write("key", data)	❌ No	Agent-to-agent context sharing
fabric.create_context("name")	❌ No	Multi-LLM session state
fabric.store_kv_cache(prefix)	❌ No	Shared system prompts, long contexts
SuperBrainMemory (LangChain)	❌ No	Chat history across restarts
enable_distributed_kv_cache()	❌ No	PyTorch/HuggingFace VRAM overflow

# ❌ Raw Client — you must free manually
ptr = client.allocate(100 * 1024 * 1024)
client.write(ptr, 0, b"data")
client.free(ptr)  # ← required!

# ✅ SharedContext — no free, ever
ctx = fabric.create_context("my-session")
ctx.write("findings", {"summary": "..."})   # stored in distributed RAM
ctx.read("findings")                        # read from anywhere

# ✅ KV Cache Pool — no free, auto-evicted
ptr = fabric.store_kv_cache(b"System prompt", model="gpt-4")
# 1000 agents → same ptr, stored once ✅

→ Full Memory Management Guide with diagrams

---

🗺️ Roadmap

Version	Milestone	Status
v0.1.0	Core Distributed RAM (Allocate/Read/Write/Free)	✅ Shipped
v0.1.1	Secure Fabric (mTLS, E2EE, Pub/Sub)	✅ Shipped
v0.2.0	Phase 3: Automated AI Memory Controller	✅ Shipped
v0.3.1	Semantic Memory (FAISS-Backed Distributed Vectors)	✅ Shipped
v0.4.0	Gossip & P2P Membership	✅ Shipped
v0.5.0	High Availability & Partition Tolerance	✅ Shipped
v0.6.0	Decentralized Observability & Metrics	✅ Shipped
v0.7.1	Tiered Architecture (L1 Shared Memory) & SHM Locality Bypass	✅ Current
v0.8.0	Raft Consensus Replication	✅ Shipped
v0.9.0	NVMe Spilling	✅ Shipped

---

🖥️ Server Requirements

This SDK connects to a SuperBrain cluster. To run one locally:

docker compose up -d   # From the main repo: github.com/anispy211/memorypool
# Dashboard: http://localhost:8080

---

📚 Documentation

• Full Documentation
• Release Guide
• GitHub Repository— BSL 1.1