FlowSync — Persistent Memory for AI Coding Agents

Give your AI agent project-wide context it never forgets.

AI for Bharat Hackathon

Team Vanta — Aahil Khan, Anoushka Awasthi, Maulik Dang, Sanyam Wadhwa

---

🧠 What is FlowSync?

FlowSync gives AI coding agents — GitHub Copilot, Cursor, Claude — persistent project memory via the Model Context Protocol (MCP).

Your agent calls log_context to record why it made a decision, and search_context to recall what the team decided weeks ago. Five MCP tools turn your agent from a stateless autocomplete into a teammate that remembers everything.

For developers who don't use an AI agent, FlowSync also auto-captures context from every git push as a built-in fallback.

---

🚨 The Problem

AI coding agents are powerful but stateless — they lose all project context between sessions. Traditional tools can't fix this:

Tool	Limitation
Git Logs	Stores what changed, not why
Documentation	Requires manual updates; goes stale quickly
AI Assistants	No persistent memory across sessions
Chat history	Fragmented, unsearchable, per-user

---

✅ How FlowSync Works

1. 🤖 Agent Logs — log_context

After completing a task, your AI agent calls log_context to record the decisions made, risks introduced, and reasoning — structured and searchable. This is FlowSync's core value: capturing the why behind code, not just the what.

2. 🔍 Agent Searches — search_context

Before starting work, your agent calls search_context with a natural-language question like "what did we decide about auth?" — and gets a grounded answer with source citations, powered by Titan Embeddings + Nova Pro RAG.

3. 📡 Auto-Capture Fallback — Git Push

For developers who don't use an AI agent, a post-push hook automatically sends diffs to Amazon Bedrock (Nova Pro) which extracts decisions, risks, tasks, and affected files. The project brain grows either way.

👥 Team Visibility*

Any team member can view the context timeline, chat with the project brain, or ask natural language questions from the web dashboard — no AI agent required.

---

✨ Key Features

• log_context MCP Tool — AI agent records decisions, risks, reasoning, and tasks after every unit of work
• search_context MCP Tool — AI agent queries project history with natural language; gets grounded, citation-backed answers
• 5 MCP Tools Total — get_project_context, get_recent_changes, search_context, log_context, get_events — works with Copilot, Cursor, Claude
• Auto-Capture Fallback — Every git push triggers AI extraction via Nova Pro; project brain grows even without an AI agent
• Strict Traceability & Source Citations — Every AI insight is linked back to its originating commit or logged context, preventing hallucinations
• Team Dashboard — Real-time timeline of decisions, risks, and tasks across all branches and contributors

---

🏗️ Architecture

CLIENT LAYER                        BACKEND LAYER (AWS)
─────────────────────────────────────────────────────────────────
VS Code Extension (VSIX)            Amazon API Gateway
  ├── Git push hook         ──►         ├── Ingestion (Lambda)
  └── MCP Tool calls                    │     └── Validates, stores to DynamoDB
                                        │         Invokes AI processing async
                                        │
MCP Server (stdio)       ◄──►           ├── AI Processing (Lambda + Nova Pro)
  ├── get_project_context               │     └── Extracts context, embeddings,
  ├── get_recent_changes                │         merge propagation
  ├── search_context                    │
  ├── log_context                       ├── MCP Handler (Lambda)
  └── get_events                        │     └── Routes 5 MCP tool calls
                                        │
Web Dashboard (Next.js)  ◄──            ├── Query (Lambda + Nova Pro)
  (hosted on S3 + CloudFront)             │     └── Natural language Q&A, RAG
                                        │
                                        └── Chat (Lambda + Nova Lite)
                                              └── Conversational interface

STORAGE
  ├── flowsync-projects  — DynamoDB (project metadata + API tokens)
  ├── flowsync-events    — DynamoDB (raw push events)
  ├── flowsync-context   — DynamoDB (AI-extracted context + embeddings)
  ├── flowsync-cache     — DynamoDB (RAG response cache, 1-hr TTL)
  └── flowsync-raw-*     — S3 (raw event archive)

---

🔄 Process Flow

1. Agent works on task  →  2. Agent calls log_context   →  3. Context stored in Project Brain
        ↑                                                           ↓
6. Team views dashboard   ←  5. RAG answers with citations  ←  4. Agent calls search_context

                  ── Fallback: git push auto-captures diffs ──

Key Actors:

• AI Agent — Primary user; logs decisions via log_context, queries via search_context
• Developer — Codes normally; pushes trigger auto-capture as fallback
• Team Lead / Member — Views dashboard, chats with project brain

---

🛠️ Tech Stack

Layer	Technology
IDE Extension	TypeScript / Node.js (VS Code)
MCP Server	TypeScript, @modelcontextprotocol/sdk, stdio transport
API & Security	Amazon API Gateway (TLS 1.3, Bearer tokens, rate limiting)
Serverless Compute	AWS Lambda (Python 3.12)
AI / LLM	Amazon Bedrock — Nova Pro (intent extraction), Nova Lite (chat/Q&A)
Embeddings	Amazon Titan Text Embeddings v1
Database	Amazon DynamoDB
Asset Storage	Amazon S3
Frontend Dashboard	Next.js 14, React 18, Tailwind CSS, shadcn/ui — hosted on AWS S3 + CloudFront

---

💰 Estimated Cost

Component	AWS Service	Monthly Cost
Intent Extraction	Amazon Bedrock (Nova Pro)	~₹420 ($5.00)
Chat & Q&A	Amazon Bedrock (Nova Lite)	~₹85 ($1.00)
Embeddings	Amazon Titan Embeddings	~₹84 ($1.00)
Compute	AWS Lambda	<₹170 ($2.00)
Storage	Amazon DynamoDB	~₹420 ($5.00)
API & Network	Amazon API Gateway	~₹265 ($3.15)
Total		~₹1,400 ($17.00) / month

~₹350 ($4.00) per developer per month (4-person team)

Why it's cost-effective:

• Zero Idle Cost — Serverless architecture means you pay ₹0 when the team isn't coding
• High ROI — Saving just a couple of hours of confusion per month makes it pay for itself
• Linear Scalability — Costs grow linearly with team activity, no sudden tier jumps

---

�️ Why This Architecture

DynamoDB over RDS

FlowSync ingests unpredictable bursts of developer events — spiky, bursty, schema-light workloads that would thrash a relational DB. DynamoDB delivers single-digit millisecond reads at any scale, and PAY_PER_REQUEST means ₹0 idle cost overnight when teams aren't coding. A JOIN-heavy RDS instance would sit idle burning reserved capacity.

Lambda over EC2 / ECS

There is no sustained load — events arrive in bursts during working hours then go silent. Lambda scales to zero between events and to hundreds of concurrent executions during a commit storm. EC2 or ECS would require capacity planning, health checks, and a baseline bill even at rest.

Model Tiering — Nova Pro → Nova Lite

us.amazon.nova-pro-v1:0 is used only for high-value, once-per-commit intent extraction where accuracy matters. us.amazon.nova-lite-v1:0 handles interactive chat and Q&A where latency matters. This splits cost and latency: Pro costs ~4× more; routing cheaper queries to Lite cuts the AI bill by ~60% for typical usage.

Async Ingestion Pipeline

The ingestion Lambda stores the raw event to DynamoDB immediately (200 ms latency), then fires invokeAsync to the AI processing Lambda. The developer's push hook gets an instant 200 OK and is never blocked waiting on Bedrock. AI processing happens in the background within seconds.

S3 Archival for Query Audit

Every raw event payload is also archived to S3 (flowsync-raw-*). This provides a full audit trail for debugging, compliance, and potential future ML training — at roughly ₹1.7/GB/month with no compute cost.

API Gateway over ALB

API Gateway provides built-in rate limiting, per-client API keys, request validation, and TLS termination — all configured with a single CDK resource. An ALB would require a separate WAF, custom auth Lambda, and manual cert rotation.

Titan Embeddings for RAG

Amazon Titan Text Embeddings v1 is natively integrated with Bedrock, requires no external vector DB, and stores 1,536-dimension embedding arrays directly in DynamoDB beside the context item. This eliminates the operational overhead of running a separate Pinecone or pgvector instance.

Scrypt over Bcrypt for Token Hashing

Project API tokens are hashed with scrypt (N=16384, r=8, p=1). Scrypt is memory-hard (not just CPU-hard like bcrypt), making GPU-based brute-force attacks ~100× more expensive for an attacker, with no observable difference to the user.

---

�🆚 FlowSync vs. Alternatives

	FlowSync	Git Logs	Documentation	AI Assistants
AI agent can log & query context	✅	❌	❌	❌
Captures why changes happen	✅	❌	❌	❌
Auto-updated (no manual work)	✅	✅	❌	❌
Persistent project memory	✅	❌	❌	❌
Natural language Q&A with citations	✅	❌	❌	Partial
MCP-native (works with Copilot/Cursor)	✅	❌	❌	❌

---

🎯 USP

• Agent-First Architecture — Built for AI agents as the primary user; MCP tools are the main interface, not an afterthought
• Persistent Memory — Unlike chat history, FlowSync stores structured, searchable project knowledge forever
• Dual Input — AI agent logging + automatic git push capture ensures no context is ever lost
• Guaranteed Accountability — Strict traceability and source citations prevent AI hallucinations about project facts
• First of its kind — The first system that gives AI coding agents persistent, project-wide memory via MCP

---

🚀 Quick Start

Prerequisites

• VS Code 1.85+
• A git repository

1. Install the Extension

Download flowsync-1.0.1.vsix from the Releases page or from the FlowSync website.

code --install-extension flowsync-1.0.1.vsix

2. Initialize Your Project

Open your repo in VS Code. Click the ⚡ FlowSync button in the status bar → Initialize Project.

FlowSync auto-detects your project name, languages, frameworks, and branch. You’ll receive a Project ID and API Token — save these and share the token with teammates.

3. Connect Your AI Agent

Your agent (Copilot, Cursor, Claude) should automatically discover 5 MCP tools: log_context, search_context, get_project_context, get_recent_changes, and get_events.

If not, add .vscode/mcp.json to your repo:

{
  "servers": {
    "flowsync": {
      "type": "stdio",
      "command": "node",
      "args": ["${workspaceFolder}/mcp-server/dist/index.js"],
      "env": {
        "FLOWSYNC_PROJECT_ID": "<your-project-id>",
        "FLOWSYNC_TOKEN": "${input:flowsync-token}"
      }
    }
  }
}

4. Start Working

• Your AI agent calls log_context after completing tasks and search_context before starting new work — automatically
• Git pushes are auto-captured as a fallback, even without an AI agent
• Open the dashboard at flowsync.site with your Project ID and Token

Try it now: Visit the dashboard and click “Try Demo Project” to explore a live project — no setup needed.

---

📚 Documentation

• Technical Deep Dive — Full architecture, implementation details, MCP server, RAG pipeline, caching, security, and more
• Performance Report — Benchmark results, latency breakdowns, extraction accuracy, and cost analysis

---

👥 Team

Name	Role
Aahil Khan	Team Leader
Anoushka Awasthi	Team Member
Maulik Dang	Team Member
Sanyam Wadhwa	Team Member

---

Built with ❤️ for the AI for Bharat Hackathon