MemoryForge AI

MemoryForge AI gives agents an operating memory: structured experience that can change the next action. The interactive demo shows an agent fail a deployment, persist the reason, retrieve it on the next run, switch strategies, and succeed.

Live demo

• App: https://memoryforge-ai-demo.ljs2546.chatgpt.site
• Video: https://youtu.be/RNgBt4bOOMI

Demo flow

1. Select Run task.
2. Watch the default deployment strategy time out.
3. Confirm that the failure and recommended correction appear in long-term memory.
4. Select Run again with memory.
5. Watch the agent retrieve the past failure, choose an asynchronous deployment, verify service health, and store the successful outcome.

Architecture

AI Agent → AWS Runtime → CockroachDB Memory → Relevant Experience → Better Action

The production architecture is designed around:

• Amazon Bedrock for agent reasoning
• AWS Lambda for serverless execution
• CockroachDB for durable, distributed memory storage
• vector retrieval for relevant past experience
• structured decision and outcome records for auditability

Hackathon

Built for the CockroachDB × AWS Hackathon.

Prerequisites

• Python 3.12
• A CockroachDB Cloud cluster with vector indexes enabled
• An AWS account with Amazon Bedrock model access and permission to deploy Lambda
• AWS CLI and AWS SAM CLI for deployment

Install

git clone https://github.com/lsh2546/MemoryForge-AI.git
cd MemoryForge-AI
python -m venv .venv

Activate the virtual environment, then install dependencies:

# macOS/Linux
source .venv/bin/activate

# Windows PowerShell
.venv\Scripts\Activate.ps1

pip install -r requirements.txt

Environment variables

Copy .env.example to .env for local work. Never commit the populated file.

Variable	Required	Purpose
DATABASE_URL	Yes	CockroachDB PostgreSQL connection string with sslmode=verify-full
AWS_REGION	Yes	Region used by the Bedrock Runtime client and Lambda
EMBEDDING_MODEL_ID	Yes	Bedrock embedding model; defaults to Titan Text Embeddings v2
MEMORYFORGE_AGENT_ID	Local demo only	Stable agent identity used by the two-run example

The production embedding contract is explicitly 1024 dimensions and matches agent_memories.embedding VECTOR(1024). The Python driver uses the packaged certifi CA bundle while retaining full TLS certificate verification in both local and Lambda environments.

Create the CockroachDB schema

Create a memoryforge database or choose an existing database, then apply the schema:

cockroach sql --url "$DATABASE_URL" --file schema.sql

The schema stores structured failures, successes, decisions, workflow state, and 1024-dimensional embeddings. memories_embedding_idx is a CockroachDB vector index used by cosine-distance recall. The Lambda database user needs only SELECT and INSERT on agent_memories.

Run the decision loop locally

Export the variables from .env, ensure your AWS identity can call the configured Bedrock model, and invoke the Lambda handler twice with the same task and agent ID:

python - <<'PY'
import json
from src.lambda_handler import handler

event = {
    "agent_id": "demo-agent",
    "task": "Deploy checkout-service v2.4 and verify production health",
}

print("RUN 1", json.loads(handler(event, None)["body"]))
print("RUN 2", json.loads(handler(event, None)["body"]))
PY

Expected causal chain:

1. Run 1 finds no relevant failure, uses synchronous_deployment, fails, and returns a new written_memory_id.
2. Run 2 recalls that ID, places it in plan.adapted_from_memory, changes the strategy to async_job_with_health_check, succeeds, and writes the outcome.

Deploy to AWS Lambda

The included template.yaml gives the function Bedrock invoke permission, keeps the CockroachDB URL server-side, and limits reserved concurrency.

sam build
sam deploy --guided

During guided deployment:

• choose the same AWS region in which Bedrock model access is available;
• provide DatabaseUrl from CockroachDB Cloud (the parameter is NoEcho);
• keep the default Titan embedding model unless another 1024-dimensional model is deliberately configured in both code and schema.

Invoke the deployed function twice with the same payload:

aws lambda invoke \
  --function-name memoryforge-ai-agent \
  --payload '{"agent_id":"demo-agent","task":"Deploy checkout-service v2.4 and verify production health"}' \
  --cli-binary-format raw-in-base64-out run-1.json

aws lambda invoke \
  --function-name memoryforge-ai-agent \
  --payload '{"agent_id":"demo-agent","task":"Deploy checkout-service v2.4 and verify production health"}' \
  --cli-binary-format raw-in-base64-out run-2.json

Compare run-1.json and run-2.json. The second response must contain the first run's memory ID in both recalled_memory_ids and plan.adapted_from_memory, plus result.status: success. This is the proof that CockroachDB recall changed the next action rather than merely displaying history.

Security notes

• Keep DATABASE_URL only in Lambda environment configuration or a secret store.
• Do not expose database credentials, Bedrock responses, or stack traces to a browser.
• Use a least-privilege CockroachDB user and a reserved concurrency limit.
• Parameterized SQL is used for every memory value; arbitrary SQL is not accepted.

Verified deployment evidence

The submitted demo was re-run against the deployed memoryforge-ai-agent Lambda in us-east-1, Amazon Bedrock Titan Text Embeddings v2, and CockroachDB Cloud. With the same agent and task payload:

• Run 1 returned failed with synchronous_deployment and wrote memory c7a7eb57-dd10-4163-8618-88c9232c2678.
• Run 2 recalled that exact ID, returned it as plan.adapted_from_memory, changed the strategy to async_job_with_health_check, and returned success.
• The automated equality and outcome check returned same_memory_changed_action: true.
• CockroachDB Cloud Managed MCP select_query independently returned the same UUID as a failure memory with decision async_job_with_health_check, outcome timeout, and confidence 0.96. The read-only reproduction script is scripts/managed_mcp_query.mjs.

The 2 minute 14 second demo video shows the application states and the live AWS invocation evidence without exposing the database URL, credentials, or environment values.