Advanced Memory Systems

Go beyond simple chat history. Implement long-term persistence, vector retrieval, and episodic memory.

Memory Types#

Short-Term

The immediate context window. Stores recent messages. Cleared when session ends.

Long-Term (Vector)

Semantic search over documents. "Read-only" knowledge base for the agent.

Episodic

Remembering past interactions with this specific user. "You mentioned you like Python last week."

Implementing Persistent Memory#

By default, `InMemoryStore` forgets everything on restart. To build a real app, use a database adapter (e.g., Redis, Postgres).

redis-memory.ts

import { MemoryStore, Message } from '@akios/sdk'
import { Redis } from 'ioredis'

export class RedisMemory implements MemoryStore {
  private client: Redis
  
  constructor(url: string) {
    this.client = new Redis(url)
  }

  async getMessages(sessionId: string): Promise<Message[]> {
    const data = await this.client.lrange(`session:${sessionId}`, 0, -1)
    return data.map(d => JSON.parse(d))
  }

  async addMessage(sessionId: string, message: Message): Promise<void> {
    await this.client.rpush(`session:${sessionId}`, JSON.stringify(message))
    // Optional: Set TTL to expire old sessions
    await this.client.expire(`session:${sessionId}`, 86400 * 7) 
  }

  async clear(sessionId: string): Promise<void> {
    await this.client.del(`session:${sessionId}`)
  }
}

Episodic Memory (The 'Profile' Pattern)#

To make agents feel personal, they need to update a "User Profile" based on conversations. This is often done with a background "Summarizer" agent.

Architecture Pattern

Don't stuff every past conversation into the context window. Instead, have a background job summarize key facts into a `user_profile` table, and inject that profile into the System Prompt.

profile-injection.ts

// 1. Fetch Profile
const profile = await db.users.find(userId)

// 2. Inject into System Prompt
const agent = new Agent({
  systemPrompt: `You are a helpful assistant.
  
  User Profile:
  - Name: ${profile.name}
  - Technical Level: ${profile.techLevel}
  - Preferred Language: ${profile.language}
  `,
  // ...
})