🧠 AURA Memory

Agentic forensic memory for deepfake detection. Every analysis is remembered as a vector, recalled by similarity, and queryable in plain language.

Built for the CockroachDB × AWS Hackathon 2026 — Build with agentic memory.

🎥 Watch the 60-second demo · 📂 Source code

---

The problem

Deepfake detectors are stateless. Every video is judged in isolation; the moment the response is returned, the evidence is gone. There is no institutional memory — no way to ask "have we seen this manipulation before?" or "which files came back with conflicting signals last week?"

AURA Memory turns a one-shot detector into a system that remembers.

What it does

• 🔬 Analyzes a video through AURA's multi-layer forensic pipeline and returns a verdict (SUSPICIOUS, CONFLICTING SIGNALS, AUTHENTIC, …) with a composite score. Layer 10 even identifies the likely generative origin (Sora / Kling / Veo).
• 🧬 Remembers every analysis: a durable detection record plus a multimodal vector embedding is written to CockroachDB.
• 🔁 Recalls similar past detections via cosine similarity — the "seen before" signal is returned inline with every new analysis.
• 🤖 Answers questions about its own memory in natural language via the Model Context Protocol (MCP) — an AI agent turns "how many detections and with which verdicts?" into SQL, runs it over a locked-down read-only connection, and answers.

Architecture

Frontend: Next.js on Vercel. Backend: Docker image on Amazon ECR, running on Amazon ECS (Express Mode). State + memory: CockroachDB Serverless on AWS.

Mermaid source

flowchart LR
    U[Client / Frontend] -->|POST /analyze| API[FastAPI backend<br/>Amazon ECS · ECR image]
    API --> P[Forensic pipeline<br/>11 layers · OpenCV · MediaPipe]
    API --> E[Amazon Bedrock<br/>Titan Multimodal Embeddings<br/>1024-dim]
    E --> DB[(CockroachDB Serverless<br/>on AWS)]
    P --> DB
    DB -->|cosine recall| API
    API -->|verdict + similar_detections| U

    A[AI Agent<br/>Claude Code / any MCP client] -->|natural language| MCP[CockroachDB MCP Server<br/>read-only]
    MCP -->|SQL| DB

Loading

Tech stack

Layer	Technology
Forensic detection	Python · FastAPI · OpenCV · MediaPipe · Transformers (11-layer pipeline)
Embeddings	Amazon Bedrock — Titan Multimodal (1024-dim)
Memory / vector store	CockroachDB Serverless (AWS) — VECTOR column, cosine similarity
API	FastAPI + OpenAPI / Swagger (/docs)
Agentic memory	Model Context Protocol — CockroachDB MCP server (read-only)
Frontend	Next.js on Vercel
Compute	Amazon ECS (Express Mode), image on ECR; AWS Lambda for ingestion

How the memory works

Each analysis produces a fixed-size 1024-dimensional embedding from Amazon Bedrock's Titan Multimodal model. Because the embedding is fixed-size, a 5-second clip and a 2-hour film cost the same storage — the video itself never enters the database, only its mathematical fingerprint. Raw files live in S3.

Detections are stored in a single table (simplified):

CREATE TABLE detections (
    id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    filename        STRING,
    verdict         STRING,          -- SUSPICIOUS | CONFLICTING SIGNALS | AUTHENTIC ...
    status          STRING,          -- done | processing | error
    composite_score FLOAT,
    embedding       VECTOR(1024),    -- Titan Multimodal fingerprint
    pdf_url         STRING,          -- forensic report (S3)
    created_at      TIMESTAMPTZ DEFAULT now()
);

Recall is a nearest-neighbour query over the vector column:

SELECT filename, verdict, embedding <-> $1 AS distance
FROM detections
ORDER BY distance
LIMIT 5;

A distance near 0 means we have analyzed something almost identical before — surfaced as similar_detections in the /analyze response.

Agentic memory via MCP ⭐

This is the heart of the agentic memory theme. The detection memory is exposed to any MCP-compatible AI client (Claude Code, Cursor, …) through the CockroachDB MCP server. Questions in natural language become SQL against the live database:

"How many detections are there and with which verdicts?" → SELECT verdict, count(*) FROM detections GROUP BY verdict

"Show me the last detection with its verdict and time."

"Are there any detections with conflicting signals? Which files?"

The agent writes the query, runs it, and even self-corrects — for example, excluding the large embedding column when it is not needed.

Security model

Two independent, demonstrable layers protect the memory:

1. Least-privilege database user — a dedicated aura_readonly SQL user with GRANT SELECT on the detections table only. It cannot write, alter, or drop anything.
2. Read-only MCP server — the CockroachDB MCP server runs with --read-only, gating every write-shaped tool at the protocol layer.

All connections are encrypted in transit (sslmode=require).

Even if the agent were compromised or prompted maliciously, it is structurally unable to modify the memory.

API

POST /analyze — multipart upload (file=@video.mp4) →

{
  "verdict": "SUSPICIOUS",
  "composite_score": 0.396,
  "status": "done",
  "similar_detections": [
    { "filename": "…", "verdict": "…", "distance": 0.02 }
  ]
}

Interactive docs: /docs (Swagger UI).

Getting started

git clone https://github.com/Artkill24/aura-memory.git
cd aura-memory
cp .env.example .env        # then fill in your values
pip install -r requirements.txt
uvicorn app.main:app --reload   # adjust to your actual entrypoint

Connect the MCP server (read-only)

claude mcp add cockroachdb -- uvx --from git+https://github.com/amineelkouhen/mcp-cockroachdb.git \
  cockroachdb-mcp-server --read-only \
  --url "postgresql://aura_readonly:***@<host>:26257/defaultdb?sslmode=require"

Then ask your MCP client: "How many detections are there and with which verdicts?"

Roadmap / hardening

• Bake the CockroachDB CA cert into the image to enable sslmode=verify-full
• Rotate the aura_readonly credential and load it from a secrets manager
• Batch-ingest a labelled deepfake/authentic corpus to enrich recall
• Expose recall thresholds via the API

Hackathon

Built for the CockroachDB × AWS Hackathon 2026 — Build with agentic memory. Submission deadline: 18 Aug 2026.

🎥 Demo video

---

Part of the AURA deepfake-detection project · ARTKILL24