Memory makes an application appear to remember earlier conversation or task details, even though the underlying LLM starts each call from a clean slate. section 1.3 section 1.4
A memory-enabled application typically follows this lifecycle:
- Capture previous user inputs and model outputs. section 1.4
- Select or retain relevant history for the next interaction. section 1.4
- Inject that history into the next prompt before calling the LLM. section 1.4
Remembering a name
After a user says, “Hi, I’m Andrea,” an isolated later question—“What’s my name?”—produces uncertainty. Resending the earlier exchange lets the application answer “Andrea.” section 1.4
Figure 1.3 In LLM relationships, every chat is a first date.
Use the “first date” metaphor to see why every direct LLM call lacks conversational continuity. section 1.4
A minimal implementation can keep a history list, append each new turn, and reinject the rendered history before the next model call. section 3.4
