Agent Memory Cheat Sheet

Lexical (BM25) score: sr
Semantic (embedding distance) score: se
Metadata filter score: sm

Combined score = w1*sr + w2*se + w3*sm
where w1 + w2 + w3 = 1

Typical: w1=0.3, w2=0.6, w3=0.1

Agent → Brain SDK → Local/Docker server
Latency: < 10ms
Scaling: Horizontal per agent

Agent → Brain Cloud API → Managed storage
Latency: < 50ms P99
Scaling: Auto-scaling, multi-region

Agent → Redis (cache) → Brain Cloud (persistent)
Latency: < 5ms (cache hit), < 50ms (miss)
Scaling: Redis cluster + managed backend

import { Brain } from "brain-ai";

const memory = new Brain({
  model: "memory",
  endpoint: "http://localhost:9090", // self-hosted server: arena + WAL + redb
  embedding: "all-MiniLM-L6-v2",
});

const memory = new Brain({
  apiKey: process.env.BRAIN_API_KEY,
  agentId: "agent-123",
  region: "us-east-1",
});

await memory.store({
  agentId: "agent-1",
  content: "User prefers async communication",
  type: "preference",
  metadata: { priority: "high" },
});

const results = await memory.retrieve({
  agentId: "agent-1",
  query: "User communication preferences",
  topK: 5,
  filter: { type: "preference" },
});

await memory.batch([
  { action: "store", payload: {...} },
  { action: "retrieve", payload: {...} },
]);