QuestorOS Memory — ICARE³ Organizational Intelligence
Inspiration
AI agents are increasingly capable of reasoning and taking action, but most still struggle with durable organizational memory. Important decisions, corrections, provenance, and context are often fragmented across conversations and systems.
QuestorOS Memory was built to give agents a persistent, governed memory layer that can preserve organizational knowledge over time while distinguishing current truth from superseded context.
The result is ICARE³: an organizational intelligence architecture designed around durable memory, retrieval, correction, revision history, explainability, and secure scoping.
What it does
QuestorOS Memory provides persistent agent memory backed by CockroachDB.
The live demonstration shows an end-to-end memory lifecycle:
- Verify CockroachDB connectivity.
- Authenticate a scoped judge credential.
- Store a synthetic organizational memory.
- Generate a 1,024-dimension embedding using Amazon Bedrock Titan.
- Persist the vector alongside the memory.
- Retrieve the memory using CockroachDB distributed vector similarity search with explainable scoring.
- Correct the memory when new evidence supersedes old information.
- Preserve immutable revision history showing both the previous and current versions.
- Delete the disposable demonstration memory and associated vector.
The demo uses only synthetic data in an isolated staging environment.
How we built it
CockroachDB acts as the persistent memory system of record.
We use CockroachDB Distributed Vector Indexing to store and retrieve semantic embeddings together with transactional memory data. This allows vector similarity retrieval and durable organizational records to live in the same distributed database.
We also use the CockroachDB Cloud Managed MCP Server as a read-only development and diagnostic interface for schema inspection and database verification.
AWS powers the agent-facing memory service:
- AWS Lambda runs the isolated Memory API.
- Amazon API Gateway exposes the staging endpoints.
- Amazon Bedrock Titan generates 1,024-dimension embeddings.
- AWS Secrets Manager protects database and judge-demo credentials.
The public hackathon repository contains a clean-room reference implementation, SQL schema, AWS SAM deployment template, configuration examples, documentation, and Apache 2.0 license without exposing proprietary QuestorOS production code or customer data.
Persistent memory and correction
A key design requirement was that a correction should not silently destroy historical context.
When information changes, QuestorOS Memory creates a new revision while preserving the superseded version. The current memory becomes the durable context used by future agents, while revision history retains who changed it, what changed, and why.
This provides agents with both current organizational knowledge and an auditable history.
Explainable retrieval
Semantic retrieval should not be a black box.
Search results expose signals used to rank the memory, including vector similarity, scope matching, keyword contribution, and the resulting score. This gives agents and developers visibility into why a memory was retrieved.
Challenges
The biggest challenge was combining agentic semantic memory with the governance expected from an organizational system.
We needed to support vector retrieval without sacrificing tenant and project isolation, correction history, provenance, explainability, or secure access.
We also deliberately separated the hackathon implementation from the proprietary QuestorOS production repository. The public repository contains only the reference implementation needed for judging, while the live demonstration runs against isolated synthetic staging data.
What we learned
CockroachDB is particularly well suited to agent memory because transactional organizational data and semantic vectors can coexist in the same distributed persistence layer.
Instead of maintaining a separate vector database and transactional database, the agent can retrieve semantic context while preserving the relational structure, scope, provenance, and revision history that organizational memory requires.
What's next
The next stage for ICARE³ is deeper autonomous memory governance: confidence-aware consolidation, contradiction detection, policy-controlled publishing, cross-agent organizational learning, and lifecycle management for memories that become stale or superseded.
The long-term goal is for AI agents to build institutional knowledge continuously without losing accountability, history, or context.
Built With
- agentic-ai
- ai-agents
- amazon-api-gateway
- amazon-bedrock
- amazon-web-services
- aws-lambda
- aws-secrets-manager
- cockroachdb
- cockroachdb-cloud
- embeddings
- mcp
- next.js
- node.js
- persistent
- rest-api
- typescript
- vector-database
- vector-search
Log in or sign up for Devpost to join the conversation.