Building TechSphereX Studio: Giving AI Agents a Memory
Have you ever found yourself correcting GitHub Copilot or Cursor for the exact same bug or architectural pattern you fixed last week?
As AI agents become more integrated into our workflows, they often lack the "institutional memory" of our specific team's codebase, security standards, and past hard-learned lessons.
That’s why I built TechSphereX Studio — an AI Experience Engine that intercepts AI actions and provides real-time, context-aware suggestions.
✨ What is TechSphereX Studio?
It is a self-learning system designed to act as a bridge between your AI agent and your team's accumulated knowledge base.
The 3-Layer Intercept Pipeline:
L1 (Read-only Filter): Instantly skips non-destructive actions like ls or cat (< 1ms).
L2 (Semantic Search): Uses vector embeddings (Qdrant) to find relevant past "experiences" (< 50ms).
L3 (LLM Anti-Noise): A local LLM (via Ollama) filters out irrelevant results to ensure high-quality suggestions (< 500ms).
🛠 The "Polyglot" Tech Stack
One of the most exciting parts of this project was working across multiple languages in a single monorepo:
🐍 Python (FastAPI + Qdrant): The core engine handling the AI pipeline and vector database.
🦀 Rust (Tauri 2.0): Powering a high-performance native desktop dashboard.
⚛️ React 19 + Vite 6: A modern UI for managing "experiences" and viewing KPIs.
📘 TypeScript: Used for the VS Code Extension, CLI tool, and shared SDK.
🏗 Architecture at a Glance
The system is designed to be modular. Whether you are using the VS Code Extension, the CLI, or the Desktop App, everything communicates with the central FastAPI engine.
graph TD
EXT["VS Code Extension"] --> API["FastAPI Core"]
CLI["CLI Tool"] --> API
DESK["Tauri Desktop"] --> API
API --> QDRANT[("Qdrant Vector DB")]
🚀 How it works with your AI Agent
Once you start the engine and configure your agent (Copilot, Cursor, Gemini, or Claude), the magic happens:
You: "Create a database connection."
Agent: Queries TechSphereX POST /api/intercept.
Engine: Returns suggestions like "⚠️ CRITICAL: Use Singleton Pattern for DB connections."
Agent: Writes the code correctly the first time. ✅
🛡 Overcoming Challenges: The Windows Defender Issue
Bundling a Python AI engine into a Windows .exe using PyInstaller came with a surprise: False Positives. Because the bootloader extracts files to %TEMP% and contains ML libraries like PyTorch, some AVs flag it. I’ve documented my mitigation steps (disabling UPX, adding PE metadata) in the README to help others facing the same struggle!
📂 Open Source & Future
This project is currently in v0.1 and is fully open-source. Upcoming features include:
Team sync for cloud experience sharing.
Native hooks for Copilot PreToolUse/PostToolUse.
JetBrains plugin support.
Check out the repo here: https://github.com/KhaiTrang1995/techspherex-studio
I'd love to hear your thoughts! How are you managing "knowledge" with your AI agents?