Working Memory

Working memory is the most transient level of the memory hierarchy, existing only for the duration of a single agent run.

Purpose

Working memory stores:

Current task context
Intermediate results
Accumulated observations
Step-to-step state

Lifecycle


Run Starts → Working Memory Created
     │
     ├── Step 1: Read/Write working memory
     ├── Step 2: Read/Write working memory
     ├── Step N: Read/Write working memory
     │
Run Ends → Working Memory Cleared

Automatic Population

The execution engine automatically populates working memory:


{
  "working_memory": {
    "input": {
      "original_request": "Offboard Sarah Chen",
      "parsed": {
        "employee": "sarah.chen@company.com",
        "action": "offboard"
      }
    },
    "observations": [
      {
        "step": 1,
        "summary": "Retrieved employee profile and access list"
      },
      {
        "step": 2,
        "summary": "Risk score calculated: 85/100 (HIGH)"
      }
    ],
    "state": {
      "risk_level": "high",
      "approved_actions": ["revoke_aws"],
      "pending_actions": ["revoke_github", "transfer_docs"]
    }
  }
}

Agent Access

Agents can read and write working memory using the memory_read and memory_write tools:

Reading


{
  "tool": "memory_read",
  "input": {
    "level": "working",
    "key": "observations"
  },
  "output": [
    {"step": 1, "summary": "Retrieved employee profile..."},
    {"step": 2, "summary": "Risk score calculated..."}
  ]
}

Writing


{
  "tool": "memory_write",
  "input": {
    "level": "working",
    "key": "state.current_focus",
    "value": "knowledge_transfer"
  }
}

Use Cases

Tracking Progress


{
  "completed_tasks": ["gather_info", "assess_risk", "revoke_aws"],
  "remaining_tasks": ["revoke_github", "transfer_docs", "send_farewell"]
}

Storing Intermediate Results


{
  "employee_data": {
    "id": "user_456",
    "name": "Sarah Chen",
    "apps": ["aws", "github", "slack"]
  },
  "risk_assessment": {
    "score": 85,
    "factors": ["competitor", "admin_access"]
  }
}

Accumulating Context


{
  "conversation_history": [
    {"role": "user", "content": "Offboard Sarah"},
    {"role": "agent", "content": "I'll process that. Risk is high..."}
  ]
}

Best Practices

Keep It Organized

Structure working memory logically:


{
  "input": { ... },
  "state": { ... },
  "results": { ... },
  "observations": [ ... ]
}

Don’t Overload

Working memory is for the current run only. For persistent storage:

Use short-term for temporary cross-run data
Use episodic for lessons learned
Use semantic for long-term facts

Clean Up

While working memory clears automatically, explicitly removing unneeded data can help with context limits:


{
  "tool": "memory_write",
  "input": {
    "level": "working",
    "key": "raw_api_response",
    "value": null
  }
}