AI Memory Systems

Introduction

Production agents need explicit memory systems — not just a longer context. Layers: working, episodic, semantic, procedural. Tools: Letta, Zep, Mem0, custom on Postgres + pgvector.

Beginner analogy: a journal + a contacts book + a skill book — different memories for different uses.

Understanding the topic

Core concepts:

• Working: current session context.
• Episodic: past sessions, retrieved by similarity.
• Semantic: structured facts and preferences.
• Procedural: tool plans the agent has learned.
• Memory writes must be deliberate — every fact tagged with source + confidence.

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, Claude Projects, Cursor's repo index, Letta (memGPT) — all production memory.

Best practices

• Tag memories with source + confidence + TTL.
• Use a periodic compaction job.

Common mistakes

• Writing memory on every turn — store fills with noise.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Four memory types.
• Q2. Why tag memories with confidence?
• Q3. Scenario: agent recalls outdated user info. Three fixes?