Improvement ideas: conversational memory with the user, Code Property Graph, CodeAct/Manus #2
Description
Code Property Graph (CPG)
is a powerful concept that creates a dynamic graph representation of your system design / architecture diagram, code that tracks dependencies in your code, searches to dead or duplicated code, helps finding bugs and do security/vulnerability and impact analysis but most importantly helps with the code comprehension: high-level of function, class and module dependencies, interact queries to your codebase. Helps LLMs to better reason.
If we can incorporate CPG principle into sparc2 copilot so LLMs could better comprehand and query code structure and its dependancies. Here are a few frameworks for this:
CodeQL
ShiftLeftSecurity/codepropertygraph
jQAssistant
joern
Fraunhofer-AISEC/cpg
FalkorDB: code-graph & backend
GitDiagram Convert any GitHub repository structure into a system design / architecture diagram
CodeAct/Manus
CodeAct relies on using Python instead of JSON function-calling, which enables solving more complex tasks in less steps.
https://github.com/langchain-ai/langgraph-codeact
Long-Term Conversation Memory
Store per project Long-Term Conversation Memory using frameworks such as Letta (ex. MemGPT) or Mem0. Mem0 stores long-time memory in triplets for retrieval (known as Knowledge Graph). Letta is also a popular and powerful framework for long-term memory.
Daytona alternative to E2B
Daytona has locally running docker containers and has more stars on GitHub. E2B is opensource but do not provide ready to use containers to run locally.
https://github.com/daytonaio/daytona
LLM Friendly Git Repo
Gitingest & Aider's RepoMap to make repo more friendly and redable by LLM.