AI that learns from past debugging experience so it never repeats the same mistakes.

SpecMem is a CLI tool that wraps your terminal commands, automatically captures errors, tracks fix attempts via git diff, and uses past debugging history to give smarter AI suggestions — or fix bugs entirely on its own.

Every AI coding tool today — Copilot, Cursor, Claude Code — is stateless. They forget everything after each session. This means:

• They repeat the same failed fixes
• They can't learn from what worked before
• Developers waste time re-explaining context
• Tokens are wasted on redundant information

SpecMem gives AI memory and experience.

Runs your command through SpecMem. If it crashes, automatically captures the error (traceback, file paths, error type), stores it in memory, and gives you an AI-powered suggestion.

After you edit your code, re-runs the command and checks if your fix worked. Captures your git diff and records it as a failed fix or successful fix. SpecMem learns what works and what doesn't.

Fully autonomous. Runs the command, detects the error, queries memory, asks Gemini to generate a fix, applies it to the file automatically, re-runs, and repeats — until the bug is fixed. Zero human input.

specmem run "python app.py"
        ↓
Executes command via subprocess
        ↓
Captures stdout / stderr / exit code
        ↓
If error detected:
    → Parses traceback automatically
    → Stores in MongoDB with embedding
    → Searches similar past bugs (vector search)
    → Injects memory into Gemini prompt
    → Returns smarter fix suggestion

"Fix login bug" → generic answer, starts from scratch

"You already tried increasing timeout — it failed.
 Root cause was async race condition → try refreshing token before retrying."

$ specmem run "python app.py"

❌ ZeroDivisionError: division by zero
📝 Episode captured!

🧠 SpecMem Suggestion:
   Add a check for an empty list before division.

$ specmem fix "python app.py"

❌ Fix Failed — recorded in memory.
   SpecMem will warn about this approach in the future.

$ specmem fix "python app.py"

✅ Fix Successful — stored in memory.
   Future errors like this will get better suggestions.

$ specmem run "python app2.py"

🧠 This parallels past issues. Add an empty list check
   before division. Do not clear the cache — that failed before.

$ specmem agent "python app3.py"

🤖 Step 1: KeyError detected → querying memory → generating fix → applying...
✅ Step 2: Command succeeded! Zero human input.

VS Code Terminal
    ↓
specmem run / fix / agent
    ↓
FastAPI Backend (13 endpoints)
    ↓
┌────────────────────────────┐
│  Gemini          MongoDB   │
│  Embeddings      Atlas     │
│  (3072-dim)      Vector    │
│                  Search    │
└────────┬───────────────────┘
         ↓
  Gemini LLM (gemini-2.5-flash)
         ↓
  Memory-Enhanced Debug Response
         ↓
  CLI Output + Streamlit Dashboard

specmem/
├── backend/
│   ├── config.py              # Environment variables
│   ├── schemas.py             # Pydantic models
│   ├── database.py            # MongoDB connection + indexes
│   ├── memory_service.py      # CRUD for memories + episodes
│   ├── error_parser.py        # Auto-parse stderr → structured error
│   ├── retrieval_service.py   # Gemini embeddings + vector search
│   ├── llm_service.py         # Gemini LLM agent + token savings
│   ├── agent_loop.py          # Autonomous fix loop
│   └── main.py                # FastAPI (13 endpoints)
├── frontend/
│   └── streamlit_app.py       # Debug Episodes dashboard
├── cli/
│   └── cli.py                 # 8 CLI commands
├── requirements.txt
└── .env.example

git clone https://github.com/VRAJ2202/LAHACK-2026.git
cd LAHACK-2026
pip install -r requirements.txt

cp .env.example .env

Fill in your .env:

MONGODB_URI=mongodb+srv://user:pass@cluster.mongodb.net/?retryWrites=true&w=majority
MONGODB_DB_NAME=specmem
GEMINI_API_KEY=your-gemini-api-key
BACKEND_URL=http://localhost:8000

Get your keys:

• MongoDB Atlas — cloud.mongodb.com (free M0 tier)
• Gemini API — aistudio.google.com

In MongoDB Atlas → Search & Vector Search → Create Search Index:

For bug_memories collection (index name: vector_index):

{
  "fields": [{
    "type": "vector",
    "path": "embedding",
    "numDimensions": 3072,
    "similarity": "cosine"
  }]
}

For debug_episodes collection (index name: episode_vector_index):

{
  "fields": [{
    "type": "vector",
    "path": "embedding",
    "numDimensions": 3072,
    "similarity": "cosine"
  }]
}

# Terminal 1 — Backend
uvicorn specmem.backend.main:app --reload

# Terminal 2 — Dashboard
streamlit run specmem/frontend/streamlit_app.py

# Terminal 3 — Use SpecMem
export GEMINI_API_KEY="your-key"
python -m specmem.cli.cli run "python app.py" --project-id my-project

1. specmem run "python app.py" → ZeroDivisionError captured
2. Developer tries bad fix → specmem fix records it as failed
3. Developer applies real fix → specmem fix records it as successful 🟢
4. specmem run "python app2.py" → Different file, same bug pattern → SpecMem immediately suggests the correct fix from app.py's memory
5. specmem agent "python app3.py" → Fully autonomous — fixes a KeyError bug with zero human input

SpecMem was built in 24 hours at LA HACK 2026.

MIT