Decision Brain
A decision-record and code-history management tool built on a graph database. It converts a repository's decision logs and git history into a graph structure, providing AI coding agents and developers with the decision context and evidence chain behind code changes.
Timeline (last 90 days)
Decision Brain is a project submitted by Peter Green at the Graph Hacks hackathon (FalkorDB × WeMakeDevs) that turns warehouse decision records and git history into a FalkorDB graph.
Decision Brain's ingestion turns git history into Commit, File, and Person nodes, and turns each decision log entry into a Decision node with a date, status, and exact source line number, linked to the commit via git blame.
Decision Brain saves every historical version of the repository's AI agent instruction files (CLAUDE.md, AGENTS.md, .cursorrules) as a fragment by date.
Decision Brain uses four workers to claim items from a shared kanban graph; the kanban itself is also a FalkorDB graph, and claiming, discovery, and handoff are all done via Cypher.
When an agent believes item A supersedes item B, another agent must each quote a passage from the original text to confirm before the handoff counts; on Invarail's 151 items, the team initiated 15 handoffs, confirmed 9, and the rest were rejected.
Decision Brain's verification mechanism requires that each quote appears verbatim in the item, each number appears within its quote, and each file name exists in git.
Decision Brain's file summary check once wrongly rejected 13 good summaries because no file contained the full string TaskStore.add, and was later changed so that each part of a qualified name must appear as an identifier.
Decision Brain uses FalkorDB's built-in algorithms algo.pageRank to answer which files carry the most reasoning, and algo.SPpaths to answer how two things are connected.
Decision Brain's query agent is built around twelve MCP tools, including why, what_breaks, disproven, who_knows, trace, origin, told, and review.
Decision Brain's answer verification requires that every cited [id] must have actually been returned by the tools in this round, every number must appear in the evidence of the cited item, and "I don't know" counts as a valid abstention only when the rest of the answer contains no assertions.
2 more timeline events
All Facts (12)
Decision Brain is a project submitted by Peter Green at the Graph Hacks hackathon (FalkorDB × WeMakeDevs) that turns warehouse decision records and git history into a FalkorDB graph.
50%UnverifiedDecision Brain's ingestion turns git history into Commit, File, and Person nodes, and turns each decision log entry into a Decision node with a date, status, and exact source line number, linked to the commit via git blame.
50%UnverifiedDecision Brain saves every historical version of the repository's AI agent instruction files (CLAUDE.md, AGENTS.md, .cursorrules) as a fragment by date.
50%UnverifiedDecision Brain uses four workers to claim items from a shared kanban graph; the kanban itself is also a FalkorDB graph, and claiming, discovery, and handoff are all done via Cypher.
50%UnverifiedWhen an agent believes item A supersedes item B, another agent must each quote a passage from the original text to confirm before the handoff counts; on Invarail's 151 items, the team initiated 15 handoffs, confirmed 9, and the rest were rejected.
50%UnverifiedDecision Brain's verification mechanism requires that each quote appears verbatim in the item, each number appears within its quote, and each file name exists in git.
50%UnverifiedDecision Brain's file summary check once wrongly rejected 13 good summaries because no file contained the full string TaskStore.add, and was later changed so that each part of a qualified name must appear as an identifier.
50%UnverifiedDecision Brain uses FalkorDB's built-in algorithms algo.pageRank to answer which files carry the most reasoning, and algo.SPpaths to answer how two things are connected.
50%UnverifiedDecision Brain's query agent is built around twelve MCP tools, including why, what_breaks, disproven, who_knows, trace, origin, told, and review.
50%UnverifiedDecision Brain's answer verification requires that every cited [id] must have actually been returned by the tools in this round, every number must appear in the evidence of the cited item, and "I don't know" counts as a valid abstention only when the rest of the answer contains no assertions.
50%UnverifiedIn a controlled experiment on 24 frozen questions (covering the three repositories Invarail, GOV.UK content-publisher, and log4brains), Decision Brain found the required evidence 20/20, versus 15/20 for the same model plus keyword search.
50%UnverifiedIn the same controlled experiment, Decision Brain scored 17/20 and 16/20 respectively on answers passing the rubric, versus 11/20 and 10/20 for keyword search; correcting false premises was 4/4 versus 2/4; correct "I don't know" was 4/4 for both.
50%