What Is Memory

Memory is the mechanism by which an AI agent retains information across conversation turns and sessions, and recalls it when needed. Where context engineering handles “what to show the model right now,” memory handles “what to keep around for later.”

Context vs. Memory

Scroll horizontally→

Aspect	Context	Memory
Timeframe	What’s visible right now	What persists across turns and sessions
Role	Supplies the immediate judgment material	Storage that can be recalled later
Example	Current instructions, the last tool result	User preferences, past task history

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Information stored in memory is still passed to the model as context whenever it’s actually used. Memory is simply a storage layer that gets loaded into context later.

Short-Term vs. Long-Term Memory

Scroll horizontally→

Type	Description	Implementation examples
Short-term memory	Working state valid only for the current session	Conversation history, the most recent tool result
Long-term memory	Information retained across sessions	User preferences, past task results, documents

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Patterns for Implementing Long-Term Memory

1. Compression via Summarization

Instead of keeping an entire conversation or log, summarize and store the important decisions, open issues, and constraints. Passing an entire long history every time crowds the context window and buries important information.

2. Recall via Vector Search

Store past information as vectors and use similarity search to recall only what’s relevant to the current task. This approach shares its underlying idea with RAG.

3. Structured Storage

User settings, profiles, and project constraints are easier to reference accurately later if stored as structured files or records rather than free-form text.

# Example user profile
Preferred answer style: Concise, bullet-point focused
Language: English
Project constraint: Never deploy to production without approval

Things to Watch Out For

• Don’t mix in stale information: Resolved issues or outdated hypotheses left in memory can lead the AI to reason from wrong assumptions
• Don’t store everything: Storing indiscriminately adds noise and buries what matters
• Keep provenance: Recording when and from which task a piece of information came makes it easier to spot when it’s gone stale

Pages in This Section

Scroll horizontally→

Page	Content
Short-Term and Long-Term Memory	Lifespan and write criteria for each type
Memory Versus RAG	How to choose between two similar mechanisms
Long-Term Memory Storage Patterns	Comparing files, structured records, vector search, and summaries

This table scrolls horizontally. Keyboard users can focus the table and use the left and right arrow keys.

Summary

• Memory retains and recalls information across conversation turns and sessions
• If context is “what to show now,” memory is “what to keep for later”
• Short-term memory is conversation history; long-term memory is implemented via summarization, vector search, or structured storage
• Storing everything indiscriminately hurts quality — organizing stale information and tracking provenance matters more