-
-
architecture_diagram
Inspiration
AI coding assistants are great at writing code, but they forget the context behind a project. Decisions, rejected approaches, experiments, and blockers get scattered across chats, GitHub, and separate AI sessions so every new conversation starts from zero, and every returning teammate has to reconstruct why something was decided.
We wanted a system that remembers not just what a team built, but why they built it that way. That became ProjectBrain the shared memory layer for humans and AI agents.
What it does
ProjectBrain captures project knowledge as structured, typed memories: decisions, rejected approaches, experiments, tasks, blockers, and activity.
Ask a question like "Why did we reject this approach?" and Moss retrieves the most relevant memories, ProjectBrain combines them with current project state, and Gemini generates a grounded answer with cited sources never a guess.
It also ingests context directly from GitHub: commits, issues, and README content all become memories.
The standout feature is conflict detection. When a proposed action contradicts an existing decision, ProjectBrain surfaces the original decision and its reasoning instead of letting that context get lost — and an AI agent repeat a mistake the team already made.
How we built it
The architecture is deliberately simple:
Frontend → FastAPI → SQLite + Moss + Gemini → Project Context → Response
FastAPI orchestrates the backend. SQLite is the structured source of truth for memories, tasks, decisions, blockers, and activity. Moss provides fast semantic retrieval over that memory. Gemini handles memory extraction from raw text, grounded answer generation, and judging ambiguous conflicts.
GitHub ingestion turns a repo's commits, issues, and README into persistent project memory.
We chose this over a complex autonomous multi-agent system on purpose. The core loop stays simple and legible:
Capture → Retrieve → Understand State → Check → Respond with Evidence
Challenges we ran into
Deciding what's actually worth remembering was the hardest problem. Storing everything creates noise the system needs to distinguish decisions from experiments from rejected approaches from current state, not just log events.
Retrieval was the second challenge. Finding information isn't enough; the system has to find the right information and hand the model enough surrounding context to reason with it.
We also had to measure, not assume, our own performance: benchmarking Moss on our actual project data showed us the real gap between cold-start and warm retrieval, and where latency was actually coming from.
Accomplishments that we're proud of
We built a working end-to-end system, not a chatbot prototype.
ProjectBrain extracts structured memories with Gemini, stores project state in SQLite, retrieves relevant context with Moss, generates grounded and cited answers, detects conflicts against prior decisions, tracks tasks and blockers, and ingests context from GitHub all connected into one practical workflow for AI-assisted development.
What we learned
Project memory is more than storage. A useful system has to understand the relationship between historical decisions and the project's current state not just retrieve facts, but reason about which ones still hold.
Separating concerns made that tractable: SQLite as the structured source of truth, Moss for semantic retrieval, Gemini for reasoning and language. That separation made the whole system easier to build, test, and measure independently.
What's next for ProjectBrain
We want ProjectBrain to become a persistent context layer that works across the tools developers and AI agents already use expanding GitHub integration, adding more data sources, improving memory freshness and lifecycle management, strengthening conflict tracking, and integrating directly with AI coding agents.
The long-term vision is simple:
GitHub stores your code. Conversations store your discussions. ProjectBrain stores the evolving understanding of what the team is actually doing.
Log in or sign up for Devpost to join the conversation.