Learning/AI Agents — Complete Guide/Lesson 10
Chapter 4·Lesson 1 of 3·10 min

Memory

Memory

Learn context, session state, durable memory, summarization, and retrieval boundaries.

Concept diagram
flowchart TB
Q[Current request] --> C[Current context]
C --> R[Memory retrieval]
R --> X[Relevant memories]
Q --> S[(Session state)]
K[(Knowledge)] --> R
C --> M[Model]

Lesson overview

Memory

Memory is information an agent can use beyond the immediate model output. The key design decision is to distinguish current context, session state, durable memory and external knowledge.

Context

Context is what the model sees for the current decision: recent messages, current task state, retrieved evidence and tool results. It is temporary and bounded.

Session state

Session state represents the current workflow: selected IDs, pending approvals, tool results, retries and status. It should be durable when a task can outlast one request.

Long-term memory

Durable memory contains facts that are useful across sessions, such as stable preferences or project identifiers. Store only information with a clear future value and retention policy.

Retrieval

Do not inject every memory into every prompt. Search for relevant memories, rank them, validate freshness and preserve provenance. Memory retrieval is a data-access step, not an authorization step.

Safety

Memory is data, not authority. A stored instruction must not override policy. Do not silently store secrets, credentials or unnecessary personal data.

Forgetting

A complete memory system supports correction, expiration and deletion. Define what happens when a user changes a preference, requests deletion or a stored fact becomes false.

The practical architecture is simple: current context for this decision, session state for this workflow and selective durable memory for reusable facts.

Learning path

Theory → Example → Code → Practice → Quiz → Challenge → Completion

0/6 done

Step 1

Theory

Memory

Memory is information an agent can use beyond the immediate model output. The key design decision is to distinguish current context, session state, durable memory and external knowledge.

Context

Context is what the model sees for the current decision: recent messages, current task state, retrieved evidence and tool results. It is temporary and bounded.

Session state

Session state represents the current workflow: selected IDs, pending approvals, tool results, retries and status. It should be durable when a task can outlast one request.

Long-term memory

Durable memory contains facts that are useful across sessions, such as stable preferences or project identifiers. Store only information with a clear future value and retention policy.

Retrieval

Do not inject every memory into every prompt. Search for relevant memories, rank them, validate freshness and preserve provenance. Memory retrieval is a data-access step, not an authorization step.

Safety

Memory is data, not authority. A stored instruction must not override policy. Do not silently store secrets, credentials or unnecessary personal data.

Forgetting

A complete memory system supports correction, expiration and deletion. Define what happens when a user changes a preference, requests deletion or a stored fact becomes false.

The practical architecture is simple: current context for this decision, session state for this workflow and selective durable memory for reusable facts.

Step 2

Example

Example: coding agent

Current repository, active branch and files changed belong to session state. A preference such as “prefer TypeScript” may be durable memory. A payment credential is neither and should never become agent memory.

Step 3

Code

Selective memory retrieval

typescript
const session = await db.sessions.get(sessionId);
const memories = await memory.search({ userId, query: currentGoal, limit: 5 });
const context = { currentGoal, session, relevantMemories: memories };

Keep memory lookup separate from authorization decisions.

Step 4

Practice

Practice

Classify ten pieces of information from your agent as current context, session state, durable memory, external knowledge or secret. Define retention and deletion behavior for each.

Step 5

Quiz

1. What is session state for?

2. Can memory override authorization?

Step 6

Challenge

Challenge

Design memory for a support agent that remembers preferences but never stores payment credentials. Include retrieval, expiration, correction and deletion rules.

Complete every stage

Work through every step in order, then the lesson will be marked complete.

Each chapter and subtopic has its own public URL under /ai-agent.