A returning customer is recognised across sessions: profile facts, preferences, and prior issues are retrieved within a token budget and traced back to the episodes that produced them. A handoff pack shows what an escalation hand-off looks like.
| Capability | How it's demonstrated |
|---|---|
| Profile facts remembered | Agent recalls name, company, plan, preferences from a prior session |
| Relevant history retrieved | Prior billing issue surfaces when the user returns with a related question |
| Token-bounded, ranked context | Context fits within max_tokens; highest-value facts and summaries survive |
| Provenance | Each fact traces back to its source episode |
| Idempotent compilation | Recompiling produces 0 new memories |
| Handoff context pack | Compact escalation brief via /v1/handoff |
For a real model in the loop, see ../support-agent-llm/.
A running Statewave server at http://localhost:8100:
docker compose up -d # from statewave-examples/pip install statewave
python support_agent.py
# or
npx tsx support_agent.ts