Within a single task, the agent carries everything it has seen so far into each new Think step. That's how pass four is smarter than pass one. But that memory has limits, and they cause specific, recognisable failures.

What Happens	Why	What It Looks Like
Early context gets crowded out	Long tasks fill the available space	Forgets the original goal by step 12
Tool outputs are large	One big result can swamp everything else	Ignores what it learned before the big result
Nothing carries between tasks	Each task starts fresh by default	Re-investigates something it solved yesterday
Summarisation loses detail	Long histories get compressed to fit	Drops the specific fact that mattered

The first row is the one to watch for in testing. An agent that performs well on short tasks and strangely on long ones is usually running out of room, and the symptom is that it forgets constraints you gave it at the start. If your agent ignores an instruction on step fifteen that it honoured on step three, that's what you're seeing.

Agentic AI and MCP for Product Managers

Module 1What 'Agentic' Actually Means

Module 2How an Agent Actually Works

Module 3MCP: What It Is and Why You Should Care

Module 4Designing the Tool Layer

Module 5Permissions and the Human in the Loop

Module 6Evaluating Agents

Module 7Cost, Latency and Unpredictability

Module 8Designing What the User Sees

Module 9Risk and Blast Radius

Module 10When an Agent Is the Wrong Answer

Module 11Final Project: Design One Real Agent End to End

Module 12Appendix: The Prompt Library

What the Agent Remembers