From Context to Agent Memory

8 min

The Assistant That Starts from Zero Every Time

In AI 102 you learned to decide what an AI should know for one task. But many real assignments are not one-off. An assistant that helps you every week faces a different problem: what should carry over from one session to the next?

Two things are happening here. First, the model only sees what is inside its context window (上下文視窗) right now; when the session ends, or the window fills up, earlier content is no longer visible to it. Second, the officer is paying a repeating cost for information that could have been kept.

• Her formatting preference (stable, worth keeping)
• The decision to wait for a revised quotation (specific, has a date, may expire)
• Today's scratch calculation (useful now, not worth keeping)
• A supplier contact's personal phone number (should probably not be stored at all)

Key takeaway: A persistent agent needs memory, but the skill is choosing what to keep, not keeping everything.

9 min

Context versus Memory

Context is what the AI can see while doing the current task. Memory (記憶) is information deliberately kept outside the current window so that it can be brought back later. Memory is not the same as context: something in memory only helps once it is selected and placed back into the context.

This course follows one storyline: Know, Remember, Retrieve, Organize, Act, Optimize. This module is the "Remember" step, and it starts by asking what deserves to be remembered at all.

Key takeaway: Context is what the AI sees now; memory is what was kept for later. Memory only helps once it is selected back into context.

10 min

Types of Memory

People use "memory" for many different things. It helps to sort by three questions: how long does it last, what kind of content is it, and whose is it? The categories below overlap on purpose; one item can be described along all three.

Key takeaway: Ask how long, what kind and whose. One piece of information can be long-term, episodic and user-specific at once.

8 min

Use Now, Keep Temporarily, Remember Long-Term, Do Not Store

When new information appears, an agent (and the person designing it) can ask one question in order: what should happen to this?

1. 1Will it still matter after this task?
   If not, use it now and let it go. Most information belongs here.
2. 2Is it stable and likely to be reused?
   A preference or a settled rule may deserve long-term memory. A one-off number does not.
3. 3Is it appropriate to keep?
   Personal data, confidential material and secrets fail this test even when they would be convenient to have.
4. 4Is it better looked up than remembered?
   Documents that change, or live values, are usually retrieved from the source when needed rather than copied into memory. Module 3 develops this.

Key takeaway: Most information is used now and let go. Keeping something long-term is a deliberate choice, and some things must not be stored at all.

8 min

Practice: Memory Decision Exercise

Below are 8 illustrative pieces of information. For each, choose one category, then discuss with a peer or the instructor. Some cases can reasonably go two ways; what matters is that you can explain your reasoning. Do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.

• Current Context: needed for this step only
• Temporary Memory: kept for the session or project, then dropped
• Long-Term Memory: kept across tasks, with a clear reason and owner
• External Knowledge: leave it at its source and retrieve it when needed
• Do Not Store: must not be kept

• Go to the Practice page: Memory Decision Exercise

Key takeaway: Deciding what happens to each piece of information is the skill. A person still owns the decision to keep something long-term.

7 min

Recap: From "What Does It Know" to "What Does It Keep"

Human, AI, Human still applies: a person defines what may be remembered, the agent applies it, and a person reviews and remains responsible for what was kept. The next module turns this into a design: how should agent memory be created, used, updated and forgotten?

• Continue to Module 2: Designing Agent Memory Systems

Key takeaway: Memory is a designed decision, not an automatic feature: use now, keep temporarily, remember long-term, or do not store.