Applied design of the MemGPT memory management strategy
Important
Read all the README.md files in Lucy/lucy. This is becoming a full port of MemGPT to a production-able library that is backend-agnositc. We want light, modular and decoupled agent rendering using the brilliant paged memory design from MemGPT, but in a commerically viable implementation. Autogen may or may not be the multi-agent orchestrator of choice, we'll see when we get there.
We have a collection of use cases that need this. So Lucy is the abstraction where we figure this out together.
- OSS Models Only Must support VLLM/Ollama service. We can use Together.ai for testing initially with an eye on VLLM-OpenAI spec compatability
- Perpetual chat with context recall if you tell a Lucy agent that "I fucking love pickles" and later ask the agent or order me a sandwich, the bot should say "extra pickles, right?"
- Agency via Python functions a clean factory for providing agents with functions they can invoke and iterate with
- Multi: User, Tenant, Agent reflect almost every SaaS application on Earth - orgs/teams/companies have many users. Agents can differentiate between different users, different teams, and the relationship between the two. note: not multiplexing conversations. More than one human in the same conversation requires a completely different kind of model training that we haven't seen yet.
- Task Solving Agents can be given work to do and/or things to accomplish with the user, like getting status on a project or booking a flight.
- grown-ass scalability can be deployed in a container, scaled horizontally, load balanced, replicated across zones etc.
since Lucy is designed to drop into an application framework with orgs, venues, and users, you need frameworks to test it in.
- fastapi_test
#TODO:django_test
each is a basic instance of that framework where we can run a suite of Lucy tests in an agnostic way.