We Built an AI Sales Assistant That Actually Remembers Customers
We ran into an interesting problem while building AI sales workflows:
Most assistants completely forget customer context between conversations.
A user explains:
- pricing concerns
- CRM integrations
- procurement blockers
…and a few days later the assistant responds like it has never seen them before.
We experimented with persistent memory using Hindsight and runtime routing using cascadeflow to see if we could improve long-running sales interactions.
One thing that surprised us was how different the responses became after repeated conversations. Early outputs were generic, but after multiple interactions the assistant started adapting to:
- customer objections
- preferred communication tone
- integration requirements
- previous meeting context
We also added runtime routing + observability:
- cheap models for extraction tasks
- stronger models for reasoning
- token tracking
- latency monitoring
- runtime traces
Still refining a lot of the system, but the behavior evolution over time has been interesting to watch.
Curious how others here are approaching long-term memory + runtime orchestration for agents.