Memory and State Management

Agents often need to remember information during or across tasks.

There are two important concepts:

State = information about the current workflow/task.

Example:

Task ID
User request
Current step
Tool results
Approval status
Errors
Next action

Memory = information retained for future interactions or tasks.

Example:

User preferences
Previous interactions
Important facts
Historical context

Consider an agent processing a loan application:

Receive application
       ↓
Validate documents
       ↓
Check credit
       ↓
Check eligibility
       ↓
Human approval
       ↓
Approve / Reject

The system must know:

Where am I?
What has already happened?
What is the next step?
What information was returned?

That is state management.

Strands Agents can be used to implement agentic systems where the agent can interact with tools and perform multi-step tasks.

Conceptually:

             Agent
               │
       ┌───────┼────────┐
       ▼       ▼        ▼
    Search   API      Database
     Tool     Tool      Tool

The agent decides which capability it needs.

AWS Agent Squad can be used for multi-agent systems.

Instead of one agent doing everything:

              User
               ↓
        Coordinator Agent
          /      |       \
         ↓       ↓        ↓
     Research  Finance   Legal
      Agent     Agent     Agent

Each specialized agent can focus on a particular responsibility.

Single agent → one agent handles the task.

Multi-agent → specialized agents collaborate/cooperate on different parts of a task.

MCP = Model Context Protocol.

In this domain, think of MCP as a standardized way for an agent/model to interact with tools and external capabilities.

Agent
  ↓
MCP Client
  ↓
MCP Server
  ↓
Tool / External System

For example:

Claude/Agent
    ↓
MCP
    ↓
GitHub tool
    ↓
Repository

Strands Agents
→ Build agentic behavior

Agent Squad
→ Coordinate multiple agents

MCP
→ Standardize agent ↔ tool interaction