Skip to content

04. Working Memory

Category: Agent Memory Module: AI Agents Prerequisites: Agent Memory Overview, Short-Term Memory, Long-Term Memory Difficulty: Intermediate

Note: Working Memory is the temporary workspace where an AI agent actively processes information while reasoning, planning, calling tools, and making decisions. Unlike Short-Term Memory, which stores conversation history, Working Memory holds only the information required for the current reasoning step and is discarded once the task is completed.


Overview

Imagine solving a math problem.

You don't memorize every intermediate calculation permanently.

Instead, you temporarily hold numbers in your mind, perform calculations, and then forget them after reaching the answer.

Humans call this Working Memory.

AI agents work in a similar way.

Consider the following request.

User


Book me the cheapest flight from
Delhi to Berlin next Friday.

The AI agent temporarily stores:

  • Source city
  • Destination city
  • Travel date
  • Available airlines
  • Flight prices
  • Cheapest option

After the booking is completed, these temporary values are discarded.

They do not belong in Long-Term Memory.

Working Memory exists only while solving the current task.


Why Working Memory Matters

Without Working Memory

User Request


Reasoning


Forget Intermediate Results


Incorrect Response

Problems

  • Cannot perform multi-step reasoning
  • Cannot combine tool outputs
  • Cannot track execution progress
  • Poor planning
  • Frequent reasoning failures

With Working Memory

             User
          AI Agent
      ┌────────┼────────┐
      ▼                 ▼
 Working Memory       LLM
 Intermediate State

Benefits

  • Multi-step reasoning
  • Tool chaining
  • Planning support
  • Temporary calculations
  • Better decision making

Working Memory vs Short-Term Memory

Many people confuse these concepts.

Working Memory Short-Term Memory
Active reasoning workspace Conversation history
Temporary calculations Previous messages
Current task only Current session
Continuously updated Appended over time
Deleted after reasoning Deleted after session

Example

Conversation

User:
My name is Mihir.

Stored in

Short-Term Memory


Reasoning

Need weather


Call API


Parse response


Compare temperature


Generate answer

Stored in

Working Memory


High-Level Architecture

                       User
                    AI Agent
      ┌──────────────────┼──────────────────┐
      ▼                  ▼                  ▼
 Working Memory    Short-Term Memory   Long-Term Memory
      │                  │                  │
      ▼                  ▼                  ▼
Task State      Conversation      User Knowledge
                        LLM

Working Memory interacts with every component during task execution.


What is Stored?

Working Memory typically stores:

Current Goal

Generate Monthly Report

Execution Plan

Collect Data


Analyze


Create Charts


Generate PDF

Tool Results

CRM API


Customer List

Intermediate Calculations

Revenue


Expenses


Profit

Temporary Variables

Current File

Current User

Current Agent

Current Step

These values exist only while executing the workflow.


Working Memory Lifecycle

Receive Task
Create Working Memory
Reason
Call Tools
Update State
Generate Response
Discard Working Memory

Unlike Long-Term Memory, nothing is permanently stored.


How Working Memory Works

User Request
Planner
Working Memory
Reason
Tool Calling
Update State
LLM
Final Response

Every reasoning step updates Working Memory until the task finishes.


Implementation

Example 1 – Core Python

A simple Working Memory implementation using a dictionary.

class WorkingMemory:

    def __init__(self):
        self.state = {}

    def set(self, key, value):
        self.state[key] = value

    def get(self, key):
        return self.state.get(key)

    def clear(self):
        self.state.clear()


memory = WorkingMemory()

memory.set("destination", "Berlin")
memory.set("travel_date", "2026-08-14")
memory.set("flight_price", 520)

print(memory.get("flight_price"))

memory.clear()

Output

520

The memory is cleared once the task is complete.


Example 2 – LangGraph

LangGraph naturally models Working Memory using State.

from typing import TypedDict

from langgraph.graph import StateGraph

class AgentState(TypedDict):
    user_query: str
    search_results: list
    final_answer: str

workflow = StateGraph(AgentState)

workflow.add_node("planner", planner)
workflow.add_node("search", search_tool)
workflow.add_node("answer", generate_answer)

Each node reads from and writes to the shared state, which acts as the agent's Working Memory throughout the workflow.


Example 3 – Production Example

Enterprise AI agents often maintain Working Memory within a workflow engine while keeping session state in Redis.

from dataclasses import dataclass

@dataclass
class WorkflowState:

    task_id: str
    current_step: str
    tool_result: dict
    completed: bool = False


state = WorkflowState(
    task_id="TASK-101",
    current_step="Search Knowledge Base",
    tool_result={}
)

print(state)

In production systems, this state is frequently persisted by workflow engines such as LangGraph Checkpointers or Temporal so that long-running workflows can resume after failures without losing execution progress.


Enterprise Use Cases

Customer Support Agent

Working Memory maintains the current support workflow.

Example:

  • Current customer issue
  • Verification status
  • Troubleshooting steps
  • API responses
  • Current ticket state
Customer


Support Agent


Working Memory


CRM API


Knowledge Base


Response

Once the issue is resolved, the Working Memory is cleared.


Software Engineering Assistant

Stores temporary information while generating code.

Examples

  • Active repository
  • Current file
  • Function being modified
  • Compilation errors
  • Test execution results

This information is only needed while solving the current programming task.


Financial Assistant

Maintains temporary calculation data.

Examples

  • Current investment portfolio
  • Calculated returns
  • Risk score
  • Recommended allocation

These values are recalculated whenever the user starts a new financial analysis.


Enterprise Workflow Agent

Stores execution state during multi-step workflows.

Example

Receive Order


Validate Customer


Check Inventory


Calculate Shipping


Generate Invoice


Notify Customer

Working Memory tracks the progress of each workflow step.


Research Agent

Maintains intermediate reasoning while collecting information.

Example

Search Web


Retrieve Documents


Extract Facts


Rank Results


Generate Summary

Retrieved information exists only while preparing the final answer.


Production Insight

Working Memory is often confused with conversation memory.

They serve different purposes.

                     AI Agent
        ┌────────────────┼────────────────┐
        ▼                ▼                ▼
 Working Memory   Short-Term Memory  Long-Term Memory
        │                │                │
        ▼                ▼                ▼
 Workflow State   Conversation      User Profile
 Intermediate     Chat History      Preferences
 Calculations

A common production architecture is:

Memory Type Recommended Technology
Working Memory LangGraph State / In-Memory Objects
Short-Term Memory Redis
Long-Term Memory PostgreSQL / MongoDB
Semantic Memory Vector Database

A good AI agent separates these responsibilities instead of storing everything in a single memory system.


Architecture Decision

Scenario Recommended Working Memory
Multi-step reasoning LangGraph State
Tool execution Workflow State Object
API orchestration In-Memory Dictionary
Long-running workflows LangGraph Checkpointer
Distributed workflows Temporal / Durable Workflow Engine

Advantages

  • Supports multi-step reasoning
  • Maintains workflow state
  • Enables tool chaining
  • Improves planning accuracy
  • Simplifies complex task execution
  • Enables autonomous decision making
  • Easily recreated when needed

Limitations

  • Temporary by design
  • Lost if the workflow fails without persistence
  • Consumes memory during execution
  • Can become inconsistent if not synchronized
  • Not suitable for storing user knowledge

Best Practices

  • Keep Working Memory task-specific.
  • Store only temporary execution state.
  • Clear Working Memory after task completion.
  • Persist workflow state for long-running tasks.
  • Avoid storing user preferences in Working Memory.
  • Minimize unnecessary state updates.
  • Monitor memory usage in complex workflows.
  • Design workflows to recover gracefully after failures.

Common Mistakes

❌ Using Working Memory to store user profiles

❌ Saving conversation history in Working Memory

❌ Never clearing temporary state

❌ Persisting intermediate calculations permanently

❌ Mixing Working Memory with Long-Term Memory

❌ Ignoring recovery for interrupted workflows


Framework Comparison

Framework Working Memory Implementation
LangChain Chain Inputs & Intermediate Variables
LangGraph Shared Graph State + Checkpointers
LlamaIndex Workflow Context & Execution State
CrewAI Task Context
OpenAI Agents SDK Run Context & Execution State

Interview Questions

What is Working Memory in an AI Agent?

How does Working Memory differ from Short-Term Memory?

Why is Working Memory required for tool calling?

What information belongs in Working Memory?

Why shouldn't user preferences be stored in Working Memory?

How does LangGraph implement Working Memory?

What happens to Working Memory after task completion?

How can long-running workflows recover Working Memory after failures?


Quick Revision

                    User
                  AI Agent
          ┌───────────┼────────────┐
          ▼           ▼            ▼
      Planner     Working Memory   Tools
               Intermediate State
                     LLM
               Final Response
              Clear Working Memory

Key Takeaways

  • Working Memory is the AI agent's temporary workspace for reasoning, planning, and task execution.
  • It stores intermediate results, execution state, and tool outputs only for the duration of the current task.
  • Unlike Short-Term Memory, it does not maintain conversation history, and unlike Long-Term Memory, it does not persist across sessions.
  • Frameworks such as LangGraph implement Working Memory through shared workflow state, enabling reliable multi-step execution and recovery.
  • Separating Working Memory from other memory types leads to more scalable, maintainable, and production-ready AI agent architectures.

References

  • LangGraph Documentation – StateGraph & Checkpointers
  • LangChain Documentation – Chains & Memory
  • LlamaIndex Documentation – Workflows
  • OpenAI Agents SDK Documentation
  • CrewAI Documentation

Next Note

05-episodic-memory.md

In the next note, we'll explore Episodic Memory, which enables AI agents to remember previous experiences, completed tasks, successes, failures, and historical interactions. You'll learn how experience-based memory helps agents improve decision-making and supports adaptive behavior in enterprise AI systems.

Enterprise AI Engineering Handbook
Building Production-Grade Enterprise AI Systems — One Chapter at a Time.