OCDB (Orbynt Cognitive Database) is an experimental 4-layer architecture designed to explore how AI agents can store temporal memory, vector embeddings, reasoning steps, and safety corrections — all inside one unified cognitive database.
Traditional SQL/NoSQL systems cannot support cognitive reasoning patterns.
OCDB proposes a new path for agent-centric memory systems.
Layer 1 — Memory Engine (Temporal Memory + TTL)
The Memory Engine provides:
- key–value storage
- JSON structured state
- time-stamped entries
- session-based memory
- automatic TTL expiration
This allows agents to maintain dynamic internal state with controlled forgetting.
Layer 2 — Vector Engine (Semantic Search)
Built with NumPy, the Vector Engine supports:
- embedding upsert
- cosine similarity search
- metadata mapping
- top-k semantic retrieval
This enables conceptual matching and retrieval-augmented behaviors.
Layer 3 — Reasoning Graph Engine
The Reasoning Graph Engine stores an agent’s internal thought process:
- nodes represent reasoning steps
- edges represent planning, validation, correction, or execution
- timestamps and metadata included
- built using NetworkX
This allows full reasoning traceability and future visualization.
Layer 4 — Safety & Correction Memory
Tracks:
- unsafe or blocked inputs
- safety violations
- corrections applied
- repeated error patterns
- model behavioral fingerprints
This layer helps agents avoid repeating unsafe mistakes across tasks.
Demo Included
The repository includes a full demonstration of:
- temporal memory in action
- vector search
- reasoning graph creation
- safety logging and correction
- unified OCDB API usage
Run:
python demo.py
GitHub Repository
https://https://github.com/abhis-byte/orbynt-database-OrbMem
Tech Stack
- Python
- NumPy
- NetworkX
- Regex-based safety pattern engine
Disclaimer
OCDB is an experimental prototype for research and conceptual exploration.
It is not a production-ready database system.
Closing Note
OCDB demonstrates how future AI-native databases may evolve:
memory that forgets, vectors that search, reasoning that forms graphs, and safety that learns over time.
Contributions, forks, and discussions are welcome.
Top comments (0)