Banana
An open-source project for memory and context orchestration in AI coding assistants. Through the MCP protocol it provides coding agents with code-relationship retrieval, architecture-decision memory and explicit rule checking, storing code AST structures in a graph database.
Timeline (last 90 days)
Banana is an open-source project that provides coding agents, via the MCP protocol, with code relationship retrieval, architectural decision memory and explicit rule checking.
Banana's three-hop multi-edge traversal latency using FalkorDB is 17.4 milliseconds, architectural decision recall latency is 1.37 milliseconds, and AST boundary check latency is 0.28 milliseconds.
Banana has already indexed the codebase into the graph database, containing 39 files, 42 classes, 84 functions, and 275 import and call relationships.
Banana uses the Edmonds-Karp maximum flow algorithm to allocate the context budget, treating code tokens as flow, dependencies as pipes, and the LLM context budget as the sink.
Banana can spin up an isolated temporary micro-graph (banana_scratch_<session_id>) within 5 milliseconds for the agent's hypothetical reasoning.
Banana's guardrail checks can inspect graph relationships directly and flag architectural violations without calling an LLM.
Banana's backend uses FalkorDB to manage a three-layer graph: banana_knowledge (code structure and decisions), banana_demand (rules and constraints), and banana_scratch_* (temporary reasoning graphs).
The Banana service is deployed on AWS Lambda and backed by FalkorDB Cloud.
The Banana project is open-sourced on GitHub and provides an online endpoint and a demo video.
All Facts (9)
Banana is an open-source project that provides coding agents, via the MCP protocol, with code relationship retrieval, architectural decision memory and explicit rule checking.
50%UnverifiedBanana's three-hop multi-edge traversal latency using FalkorDB is 17.4 milliseconds, architectural decision recall latency is 1.37 milliseconds, and AST boundary check latency is 0.28 milliseconds.
50%UnverifiedBanana has already indexed the codebase into the graph database, containing 39 files, 42 classes, 84 functions, and 275 import and call relationships.
50%UnverifiedBanana uses the Edmonds-Karp maximum flow algorithm to allocate the context budget, treating code tokens as flow, dependencies as pipes, and the LLM context budget as the sink.
50%UnverifiedBanana can spin up an isolated temporary micro-graph (banana_scratch_<session_id>) within 5 milliseconds for the agent's hypothetical reasoning.
50%UnverifiedBanana's guardrail checks can inspect graph relationships directly and flag architectural violations without calling an LLM.
50%UnverifiedBanana's backend uses FalkorDB to manage a three-layer graph: banana_knowledge (code structure and decisions), banana_demand (rules and constraints), and banana_scratch_* (temporary reasoning graphs).
50%UnverifiedThe Banana service is deployed on AWS Lambda and backed by FalkorDB Cloud.
50%UnverifiedThe Banana project is open-sourced on GitHub and provides an online endpoint and a demo video.
50%