Where MorrowCache wins
Repeated and paraphrased questions stop paying twice. Agents that retry the same job can HIT on final text instead of another full completion. Tool-call steps still run, we never store those.
All comparisons
The baseline
vs always calling the model
The default: every chat request hits the upstream model. Simple, fresh, and expensive on repeats.
Request 1
explain mutexes
Request 2
explain mutexes
Same question, paid twice
Model calledWhere they win
Always-call is simplest and always fresh. No cache bugs, no wrong hits, no proxy to run. If volume is low or every answer must be live, just call the model.
Verdict
Use always-call until repeats hurt. Then point baseURL at Morrow and keep fail-open on.
Other comparisons