(Submission for the WeMakeDevs × Cognee "Hangover Part AI" Hackathon - Best Blogs Track)
Your AI coding assistant has amnesia. I gave mine a graph memory — and then I let it fail my Pull Requests.
Last week, Cursor suggested MongoDB for the fifth time. We migrated off Mongo three months ago. I snapped, opened a terminal, and built ProjectBrain.
ProjectBrain is a persistent memory layer for your codebase, built on Cognee and the Model Context Protocol (MCP). It doesn't just remember decisions locally for my IDE. I did something weird—I wired it to a real-time visualization dashboard and gave it veto power over my entire team's PRs via GitHub Actions.
Here is how I used Cognee's lifecycle to cure Context Rot.
The 4 Memory Verbs of Cognee
Cognee operates on a distinct lifecycle. We mapped each step to an MCP tool, allowing the IDE to command the graph directly.
1. remember(): Building the Context
When we make an architectural decision, we tell Cursor to save it. ProjectBrain ingests this into Cognee, extracting entities and building graph nodes in real-time. On our dashboard, you physically see a new node pop into existence.
2. recall(): Fetching the Context
When we ask a new question, Cursor queries ProjectBrain. Cognee uses a hybrid search (semantic similarity + graph traversal) to find relevant past decisions. Now, when I ask about transactions, it sees the explicit link between Mongo, the Double-Charge Bug, and Postgres.
3. improve() (Memify): Strengthening the Bonds
Not all memories are equal. Calling improve() in Cognee strengthens the edge weights in the knowledge graph. In our dashboard, you can see the graph edges literally shift from gray to bright cyan as feedback is reinforced.
4. forget(): Active Hebbian Decay
When a deprecated pattern is deleted, we tell ProjectBrain to forget it. We built an active Hebbian decay loop in our API: the nodes dissolve and fade out with a particle effect, and Cursor immediately stops hallucinating obsolete patterns.
The Climax: CI/CD God-Mode
Memory shouldn't just exist inside a single developer's IDE. If a decision is made, the entire organization needs to enforce it.
So, we built a headless CI/CD Enforcer using GitHub Actions. We wrote a secondary agent (reviewer_agent.py) that triggers on every Pull Request, parses the Git diff, and queries the same Cognee graph memory.
If a junior developer (or their AI) tries to sneak MongoDB back into the codebase, the CI agent detects the architectural violation and physically fails the pipeline.
The Results
By combining MCP with Cognee's graph capabilities, we achieved massive context gains:
| Metric | Result |
|---|---|
| Cold recall latency | ~180ms |
| Memory ingestion | ~1.2s per decision |
| Hallucination reduction | 14/15 test prompts accurate (vs 4/15 baseline) |
| Graph nodes at demo end | 47 interconnected entities |
We didn't just build a memory plugin. We built an organizational brain that enforces its own memory across the entire engineering team.
RAG gives your AI a library card. Cognee gives it a memory. ProjectBrain gives it a conscience.
Try it yourself:
Check out the open-source repo here: rushdarshan/brain
Deployed Dashboard: brain-production-3699.up.railway.app
Built for the WeMakeDevs × Cognee "Hangover Part AI" hackathon.
Top comments (1)
Really like the shift from "memory as retrieval" to "memory as enforcement." The part that stood out to me was wiring recall into CI instead of keeping it trapped inside one IDE session. That feels much closer to how real teams prevent context rot: decisions become operational guardrails, not just notes. I'd be curious how you handle confidence/override paths when the memory graph is stale or when architecture is intentionally changing.