Backend-agnostic memory-layer SDK — pluggable providers, local embeddings, storage adapters, semantic search.

Docs: docs.atomicstrata.ai/sdk

AtomicMemory Core currently reaches cost-Pareto SOTA on BEAM-100K, BEAM-1M, and LoCoMo10, with BEAM-10M parity against the strongest published Mem0-new result. The SDK is the typed application surface for building on that memory layer.

What this package provides

• AtomicMemoryClient — primary public surface. Aggregates the memory and storage namespaces: client.memory.search(...) and client.storage.put(...).
• Provider interface + registry — implement MemoryProvider to plug in any backend.
• AtomicMemoryProvider — HTTP adapter for atomicmemory-core.
• Mem0Provider — HTTP adapter for Mem0 (OSS or hosted).
• StorageManager — KV / cache adapters under the ./kv-cache subpath (IndexedDB, in-memory).
• EmbeddingGenerator — local embedding generation via transformers.js.
• SemanticSearch — cosine-similarity search primitives.
• Error types (AtomicMemoryError, StorageError, SearchError, plus storage typed errors like ArtifactInUseError, PointerContentNotManagedError) and a minimal event emitter.

Server-side only in v1. The direct storage API uses a shared bearer credential and must run inside a trusted process (a Node server, ops tooling, or the webapp-sdk proxy). Browser bundles must NOT instantiate AtomicMemoryClient directly.

Installation

pnpm add @atomicmemory/sdk

Also works with npm install / yarn add.

Quick start

Prerequisite: start atomicmemory-core first. The full SDK walkthrough is in the SDK Quickstart.

import { AtomicMemoryClient } from '@atomicmemory/sdk';

const client = new AtomicMemoryClient({
  apiUrl: 'http://localhost:3050',
  apiKey: process.env.ATOMICMEMORY_API_KEY!,
  userId: 'demo-user',
  memory: {
    providers: {
      atomicmemory: { apiUrl: 'http://localhost:3050' },
    },
  },
});
await client.memory.initialize();

// Memory namespace.
await client.memory.ingest({
  mode: 'messages',
  messages: [{ role: 'user', content: 'I prefer aisle seats.' }],
  scope: { user: 'demo-user' },
});
const results = await client.memory.search({
  query: 'seat preference',
  scope: { user: 'demo-user' },
});

// Storage namespace.
const artifact = await client.storage.put({
  mode: 'pointer',
  uri: 'https://example.com/file.pdf',
  contentType: 'application/pdf',
});
console.log(artifact.artifactId);

Applications that only need memory operations can still use MemoryClient directly. New integrations should prefer the namespaced AtomicMemoryClient.memory surface.

Providers

AtomicMemory (recommended for self-hosted)

const memory = new MemoryClient({
  providers: {
    atomicmemory: {
      apiUrl: 'http://localhost:3050',
      apiKey: process.env.ATOMICMEMORY_API_KEY,
      timeout: 30_000,
    },
  },
});

Mem0

const memory = new MemoryClient({
  providers: {
    mem0: {
      apiUrl: 'http://localhost:8888',
      apiStyle: 'oss',
    },
  },
});

Subpath exports

• @atomicmemory/sdk/browser — browser-safe entry: MemoryClient + memory types/adapters, without the root bundle's storage/embedding/search surface
• @atomicmemory/sdk/storage — storage artifact client + types (ConcreteStorageClient, StorageClient, StoredArtifact, error classes)
• @atomicmemory/sdk/kv-cache — KV / cache adapters (IndexedDB, in-memory) used internally by the embedding cache
• @atomicmemory/sdk/embedding — embedding generator
• @atomicmemory/sdk/search — semantic search primitives
• @atomicmemory/sdk/utils — shared utilities
• @atomicmemory/sdk/core — error types + events
• @atomicmemory/sdk/memory — memory types, provider interface, provider adapters

Development

pnpm install
pnpm build
pnpm test
pnpm typecheck

Refreshing mapper test fixtures

The AtomicMemoryProvider mappers are guarded by a record/replay test suite that runs against captured atomicmemory-core HTTP responses. When core's wire shape changes, refresh the fixtures:

# In sibling atomicmemory-core checkout: ensure .env has a real
# OPENAI_API_KEY (or LLM_PROVIDER=ollama), then:
docker compose up -d --build

# Back in this repo:
pnpm fixtures:capture

See src/memory/atomicmemory-provider/__tests__/fixtures/README.md for the full procedure and what gets normalized at capture time.

Contributing

Issues and PRs welcome.

License

Apache-2.0 © AtomicMemory