AI & Logic Systems develops architectures in which generative flexibility is paired with rules, provenance, verification, and memory discipline. The division is the technical home for ApolloGPT and related open research into dependable agent systems.
Questions worth building a program around.
Long-running AI work needs continuity, inspectable decisions and a reliable relationship between reasoning and action. AI & Logic Systems is the technical home for ApolloGPT: a proposed local-first agent operating architecture connecting models, tools, memory, files and human oversight. The research question is how to make a complex workflow dependable across many steps, especially when evidence changes, tools fail or context must move between sessions.
From an idea to inspectable work.
Separate source records from summaries and distinguish durable memory from temporary context. Each tool operation should carry an explicit authority boundary and leave an interpretable record. Evaluation should cover entire tasks: recovery after interruption, preservation of user decisions, handling of contradictory evidence, and correct reporting of unfinished work. Symbolic constraints and verification can complement generative flexibility, but every extra layer needs a demonstrated purpose. Representative tasks and failure logs should guide architecture changes.
How can symbolic constraints make generative systems more inspectable?
What memory architecture preserves context without quietly rewriting history?
Which evaluations measure reliability in long-running, tool-using agents?
- 1
Publish the ApolloGPT architecture note
- 2
Release representative evaluation tasks
- 3
Open selected components for technical review
Where a contribution can matter.
Priority contributions include realistic evaluation tasks, memory-schema review, privacy analysis, integration testing and reproducible implementation examples. Research demonstrations should describe exactly what operates today and what remains architectural work.
Program support can help fund documentation, computation, reproducible tools and qualified review. Explore research partnerships or support the institute.