What is the difference between prompt, context, and memory engineering?

In detail

1. Prompt engineering: the wording

This is the instruction you hand the model. The phrasing, the examples, the format you ask for. It is the most visible layer and the one everyone started with, which is exactly why it is the least differentiated today. The models got good enough that clever wording stopped being a moat. Useful, still necessary, but it is table stakes now.

2. Context engineering: the workspace

This is everything you load into the model’s context window for one task: the relevant documents, the recent history, the tools it can call. Aishwarya Srinivasan describes the shift from prompt to context engineering in episode 42 at 2:20 as the moment you stop treating the model as a magic box and start treating the thing around it as a software system. Anthropic’s engineering team puts it the same way: context engineering is “the natural progression of prompt engineering,” the work of curating everything that lands in the context window, not just the prompt. Jerry Liu goes further: his company is betting that high-quality context stays durable even as the agents themselves improve. Summarizing what is in that window to make it fit is still context engineering. Anthropic’s guide to effective context engineering puts the goal as finding the smallest set of high-signal tokens that makes the outcome you want most likely, and calls the summarizing move compaction: summarize a conversation near the window’s limit and start a fresh window from the summary. The context engineering entry has more.

3. Memory engineering: the persistence

This is what lives outside the context window. Richmond Alake draws the cleanest line on it in episode 53 at 36:33: if you are summarizing the context window to make it fit, that is context engineering. The moment you move information out to an external store and give the agent a way to retrieve it later by ID, you are doing memory engineering. Now you own retrieval speed, decay, organization, and storage. It is its own discipline, sitting at the intersection of data engineering and agent engineering. The same Anthropic guide describes the simplest version, which it calls structured note-taking or agentic memory: the agent regularly writes notes to a store outside the context window and pulls them back in later, the way a coding agent keeps a to-do list. Agent memory covers the larger designs.

The one-line test

Shrinking what is in the window right now? Context engineering. Moving something to an external store to pull back later? Memory engineering. Writing the instruction itself? Prompt engineering.

Why it matters

As frontier models converge, the work moves down this list. Prompt wording matters less as the models improve. What you feed the model, and what it remembers, is the part you actually control. Richmond’s framing at 30:48: memory is a battleground where a single developer can compete with a research lab, unlike training a frontier model.

Disclosure: Episode 42 aired while Chain of Thought was Galileo’s podcast and Conor worked there. Episode 53 was sponsored by Galileo, the Season 3 presenting sponsor (thanks at 0:28).