PowerMem TypeScript SDK

Pure TypeScript memory system for AI agents — a full port of PowerMem.

PowerMem combines vector-based semantic search with LLM-driven intelligent memory extraction, Ebbinghaus time-decay, and multi-tenant isolation. This package is a complete TypeScript reimplementation — zero Python dependency.

Features

• Pure TypeScript — no Python, no subprocess, no server needed
• Dual storage backend — SQLite (default) or SeekDB (HNSW-indexed, OceanBase compatible)
• LLM-driven intelligent add — extracts facts, deduplicates, merges with existing memories
• Semantic search — cosine similarity over real embedding vectors
• Pluggable providers — any LangChain.js embedding/LLM provider (OpenAI, Qwen, Ollama, Anthropic, etc.)
• Multi-tenant — userId/agentId/runId isolation on all operations
• CLI — pmem command for memory CRUD, stats, backup/restore, interactive shell
• Dashboard — Express-based web dashboard with stats, charts, memory management
• Agent memory — scope/permission/collaboration management for multi-agent systems
• User profiles — profile extraction, profile-aware search, query rewriting
• Ebbinghaus decay — time-based memory score adjustment with access reinforcement
• 504 tests — unit, integration, regression, e2e (Ollama), SeekDB (macOS + Linux), BDD

Quick Start

npm install powermem-ts

Simplest usage (env vars)

# .env
EMBEDDING_PROVIDER=openai
EMBEDDING_API_KEY=sk-...
EMBEDDING_MODEL=text-embedding-3-small
LLM_PROVIDER=openai
LLM_API_KEY=sk-...
LLM_MODEL=gpt-4o-mini

import { Memory } from 'powermem-ts';

const memory = await Memory.create(); // reads from .env

await memory.add('User likes coffee', { userId: 'user1' });
const results = await memory.search('preferences', { userId: 'user1' });
console.log(results.results);

await memory.close();

Explicit LangChain instances

import { Memory } from 'powermem-ts';
import { OpenAIEmbeddings, ChatOpenAI } from '@langchain/openai';

const memory = await Memory.create({
  embeddings: new OpenAIEmbeddings({ model: 'text-embedding-3-small' }),
  llm: new ChatOpenAI({ model: 'gpt-4o-mini' }),
});

SeekDB backend (HNSW-indexed)

const memory = await Memory.create({
  embeddings: myEmbeddings,
  seekdb: { path: './seekdb_data', dimension: 1536 },
});

Connect to existing PowerMem server

const memory = await Memory.create({
  serverUrl: 'http://127.0.0.1:19527',
});

CLI

npx pmem memory add "User likes coffee" --user-id user1
npx pmem memory search "preferences" --user-id user1
npx pmem memory list --user-id user1 --sort created_at --order desc
npx pmem stats --json
npx pmem config show
npx pmem manage backup --output backup.json
npx pmem shell  # Interactive REPL

API

Memory facade

Method	Description
Memory.create(options?)	Create instance (NativeProvider default, HttpProvider with serverUrl)
add(content, options?)	Add memory (optional LLM fact extraction with infer: true)
search(query, options?)	Semantic search with scores, threshold, limit
get(id)	Get by ID
update(id, content)	Update content (auto re-embeds)
delete(id)	Delete by ID
getAll(options?)	List with pagination, sorting, filtering
count(options?)	Count with optional filters
addBatch(items, options?)	Batch add
deleteAll(options?)	Bulk delete with filters
reset()	Clear all
close()	Release resources

Configuration

Option	Description
embeddings	LangChain Embeddings instance
llm	LangChain BaseChatModel instance
dbPath	SQLite file path (default: ~/.powermem/memories.db)
seekdb	SeekDB config: { path, database?, dimension?, distance? }
serverUrl	Connect to existing server (HttpProvider mode)
customFactExtractionPrompt	Override LLM fact extraction prompt
customUpdateMemoryPrompt	Override LLM action decision prompt
fallbackToSimpleAdd	Fall back to simple add when LLM extracts nothing
reranker	Async function to re-score search results
enableDecay	Enable Ebbinghaus time-based score decay
decayWeight	Decay influence weight (0-1, default 0.3)

Architecture

src/
├── core/           Memory facade, NativeProvider, HttpProvider, Inferrer
├── storage/        VectorStore interface, SQLiteStore, SeekDBStore, factory, adapter
├── integrations/   Embeddings, LLM, Rerank — base interfaces + factories
├── intelligence/   Ebbinghaus decay, MemoryOptimizer, ImportanceEvaluator
├── prompts/        All LLM prompt templates (fact extraction, update, importance, graph)
├── utils/          Cosine search, Snowflake IDs, filter parser, stats, IO
├── cli/            Commander.js CLI (config, memory, stats, manage, shell)
├── dashboard/      Express server + HTML dashboard
├── agent/          AgentMemory, scope/permission/collaboration management
└── user-memory/    UserMemory, query rewrite, SQLite profile storage

89 source files, 40 test files. See docs/architecture.md for details.

Testing

npm test          # 370 unit/integration/regression tests
npm run test:e2e  # 21 e2e tests (requires Ollama + nomic-embed-text + qwen2.5:0.5b)
npm run test:seekdb # 63 SeekDB tests (requires @seekdb/js-bindings)
npx vitest run tests/bdd/  # 50 BDD tests (CLI + dashboard UI)

CI runs 7 jobs: Node 18/20/22 unit tests, SeekDB on macOS ARM64, SeekDB on Linux x64, build verification, e2e with Ollama.

Dependencies

Runtime: better-sqlite3, @langchain/core, commander, express, zod, dotenv

Peer (install what you need): @langchain/openai, @langchain/anthropic, @langchain/ollama, seekdb

Related

• PowerMem — Original Python implementation
• SeekDB — OceanBase embedded vector database

License

Apache License 2.0