The concept
A language model keeps nothing between calls. Every request starts fresh. Anything it "remembers" is something the app feeds back into the context window (Day 20).
Agents (Day 35) usually juggle three kinds of memory:
- Short-term: the current conversation, sitting in the context window. Gone when the session ends.
- Long-term: facts the app saves to a database and retrieves later, often with semantic search (Day 29).
- State: where a task stands right now. A checklist or a status.
Like a good HR coordinator: they remember today's chat, keep notes in your file, and track your checklist.
Walkthrough: an HR onboarding assistant
Priya joins a bank as a business analyst. An assistant helps her for two weeks.
Short-term (day 1):
- Priya asks "Where do I pick up my badge?" then "What time does that desk open?"
- "That desk" only makes sense because the first question is still in the context window.
Long-term (saved facts):
- Start date: Monday. Team: Payments. Manager: Daniel.
- Laptop: shipped, tracking number saved.
State (a checklist the app tracks, not the model):
- Say the plan has 10 tasks. By day 5, 6/10 are done: badge, laptop, payroll form, conduct training, team intro, system access.
Day 8, a brand-new session:
- Priya asks "What's left for me?"
- The app loads her saved facts and the checklist into the prompt. The model answers: "4 tasks left. Benefits enrollment is due Friday. Want the link?"
The model didn't remember. The app did, and handed it the right notes.
What NOT to remember
- Store only what the job needs.
- Sensitive details (salary, health info shared for a benefits question, ID numbers) need strict rules, or shouldn't be saved at all.
- Let people see and delete what's stored about them.
- Expire it. Onboarding notes shouldn't follow Priya for years.
Where you see it at work
- Banking: an assistant that knows a case is already open, so customers don't repeat themselves.
- Insurance: claim status kept as state, so work resumes at "awaiting photos".
- Healthcare: tight limits on what patient information an assistant may keep.
- Product management: what to remember is a product and privacy decision.
In one sentence
Models don't remember; apps give agents memory by saving the right facts and task state and feeding them back in, and deciding what not to store matters just as much.