Memory Systems

Introduction

Agents need memory to be useful beyond one turn. Three kinds: working (current task), episodic (past sessions), semantic (facts and skills). Vector DBs power most long-term memory.

Beginner analogy: humans have short-term memory (what you're doing now), long-term memory (your childhood) and skills (riding a bike). Agents need all three.

Understanding the topic

Core concepts:

• Working memory: the current context window.
• Episodic: stored past interactions, retrieved by similarity.
• Semantic: structured facts (user preferences, business rules).
• Procedural: learned tool skills (cached plans).
• Memory writes need eviction policies — don't store everything.

Syntax reference

Visual workflow / architecture:

┌────────────────┐
│  Working Memory│  (current convo)
└──────┬─────────┘
       ▼
┌────────────────┐
│  Episodic Mem  │  (past sessions)
└──────┬─────────┘
       ▼
┌────────────────┐
│  Semantic Mem  │  (facts · vector DB)
└──────┬─────────┘
       ▼
   Reasoner ← retrieves

Real-world use

ChatGPT 'Memory' feature, Claude Projects knowledge, Cursor's repo-wide context, Letta (memGPT), Zep — all production memory systems.

Best practices

• Write fewer, higher-quality memories.
• Use TTLs and confidence scores.

Common mistakes

• Storing every message verbatim — memory store fills with noise.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Four types of agent memory.
• Q2. Why is eviction important?
• Q3. Scenario: your agent remembers a user fact incorrectly. Where's the bug?