Amadeus exists to become a credible resume project for AI agent development. It should be grounded in Akashic's real design, but the near-term target is not to complete a long training track. The near-term target is to build a runnable, verifiable agent system whose core claims can survive interview follow-up.
- Real LLM turns can run through the passive runtime and persist useful session state.
- The memory system supports readable long-term memory, vector/keyword retrieval, source references, correction, forgetting, and traceable evidence.
- Evaluation exists as a product capability: cases, runner, report, and regression checks for memory, context, tools, and proactive decisions.
- Telegram-first outbound and proactive behavior can be demonstrated end to end.
- The user can explain the main code paths, tradeoffs, failure modes, and verification evidence behind each resume claim.
- Align documentation and working prompts with interview delivery.
- Confirm and harden the existing passive runtime, context, tool loop, phase, plugin, and memory behavior.
- Build productized Evaluation before expanding proactive behavior.
- Add Telegram outbound and scheduler support.
- Implement a minimal but real ProactiveLoop.
- Add DriftRunner only after the proactive foundation is demonstrable.
- Akashic is a reference implementation, not a base class.
- Amadeus should preserve clean boundaries:
MemoryEngine, public runtime behavior, eval runner, outbound adapter, scheduler, proactive pipeline. - Do not add impressive resume terms unless the repository has code and verification evidence to support them.
- Do not use fake production mechanisms to make a demo look complete. Deterministic fakes are acceptable in tests and eval fixtures.
- Generating formal study artifacts.
- Completing every Akashic subsystem before interview delivery.
- Adding QQ Bot, full MCP integration, dashboard UI, or advanced Drift behavior before Telegram, Evaluation, and ProactiveLoop have a stable vertical slice.