agentic-memory — AI memory — conversation + knowledge

Memory that
manages itself

Conversation context + knowledge retrieval. Auto-trimming, persistence, and vector search. Zero dependencies.

4KB

gzip

dependencies

layers

GitHub Try it ↓

$ npm i agentic-memory copy

Try it live

Send messages and watch auto-trimming in action. Set max to 500 tokens, then flood it.

Tokens

Messages

Turns

—

Summary

Simple by design

Manages the messages array so you don't have to.

Basic usage
import { createMemory } from 'agentic-memory'

const mem = createMemory({
  maxTokens: 8000,
  systemPrompt: 'You are a helpful assistant.',
})

await mem.user('What is quantum computing?')
await mem.assistant('Quantum computing uses qubits...')

mem.messages()  // → ready for any LLM API

With agentic-core
import { ask } from 'agentic-core'
import { createMemory } from 'agentic-memory'

const mem = createMemory({ maxTokens: 8000 })

async function chat(prompt) {
  await mem.user(prompt)
  const result = await ask(prompt, {
    history: mem.history(),  // ← multi-turn context
    apiKey: '...',
  })
  await mem.assistant(result.answer)
  return result.answer
}

await chat('Explain CRISPR')
await chat('How does it compare to older methods?')  // has context!

Persistence
// Browser
createMemory({ storage: 'localStorage:my-chat' })

// Node.js
createMemory({ storage: 'file:./chats/session.json' })

// Custom adapter
createMemory({ storage: { save(d) { ... }, load() { ... }, clear() { ... } } })

Everything you need

🔄 Auto-trimming

Sliding window drops oldest pairs. Or summarize mode compresses old context. You choose.

📊 Token estimation

CJK-aware token counter. No tiktoken dependency (40MB). ~90% accurate, good enough for budget management.

💾 Persistence

localStorage, file, or custom adapter. Conversations survive page reloads and server restarts.

🔀 Fork

Branch a conversation to explore different paths. Like git branch for chat context.

📤 Export / Import

Serialize full state to JSON. Move conversations between clients, backup, or analyze.

🏗️ Multi-conversation

createManager() handles multiple chats. Get by ID, list, delete. Each with its own storage.