Putting the last few conversations into a prompt does not give an AI long-term memory.
It is closer to dumping every file on a desk and hoping the model finds the important pages. That can work for a short demo. Over time, tokens become expensive, context becomes noisy, and critical information gets buried.
Cogweald separates its story bible into memories with different lifetimes:
- Current character state
- Active plot threads
- Recently closed threads
- Long-term facts
- Summaries of the latest chapters
Each category has a different purpose and context window. Recent summaries provide local continuity. Facts preserve information that must not drift. Threads show the model what is still unfinished.
Context engineering is not about putting more information into the prompt. It is about putting the right information there at the right time.
If your agent must read the entire database before every decision, the problem may not be the model. Your memory model may need a better design.
Top comments (0)