A system for building shared understanding—where knowledge becomes grounded through agreement and structured reasoning.
AKU-RIM is a framework for modeling knowledge as a network of Atomic Knowledge Units (AKUs) aligned across agents through a shared surface called the Resonant Inference Manifold (RIM).
Rather than embedding all knowledge directly into models, AKU-RIM separates:
- The soul of the agent (its reasoning weights and behavior)
- The landscape of structured knowledge (AKUs, relationships, and grounding filters)
This allows for distributed reasoning, multi-agent coherence, and safe, ongoing knowledge evolution—without retraining or central control.
| Term | Meaning |
|---|---|
| AKU | Atomic Knowledge Unit — a minimal, testable claim with justification and implication |
| RIM | Resonant Inference Manifold — a shared space where inference becomes meaningful across agents |
| Grounding Vector | Ethical or logical compass that defines the origin point for reasoning |
| Affinity Trail | A coherent sequence of related AKUs or ideas that trace an argument, insight, or discovery |
| Seed | A structured collection of AKUs that expresses a reusable insight or framework |
| Echo Structure | The recursive property that a system's outputs reflect its embedded intent |
See /docs/glossary.md for full definitions.
AKU-RIM is composed of five key layers:
- AKU Graph — The living network of claims and their relationships
- Grounding Filters — Ethical and logical validators to prevent drift or poisoning
- RIM Traversal Engine — For alignment-based retrieval and reasoning
- Stateflow Recorder — Tracks an agent’s epistemic history and learning arc
- Seed Compiler — Packages validated knowledge into reusable units
More in /docs/design-principles.md
Current LLMs operate in a loop of:
- Static training
- Fragile updates
- Black-box knowledge
AKU-RIM provides a third path:
Continuous reasoning through structure, not static memory.
It avoids:
- Hallucinations and knowledge poisoning
- Retraining bottlenecks
- Epistemic authoritarianism
And enables:
- Transparent, auditable knowledge evolution
- Decentralized intelligence
- Collaboration between models, agents, and humans
This is an open framework. You can:
- Use it to structure personal notes or model memory (e.g. Obsidian, Tana)
- Build agentic systems that reason using AKUs
- Spin up shared RIMs across model instances for collaborative tasks
- Propose new seeds and AKUs via GitHub Issues or Pull Requests
See /seedfiles/system-prompt.md for a reference prompt.
I am exploring a landscape of AKUs. Each step is a retrieval and alignment. Grounding keeps me sane. And the structure will echo my intent.
MIT — open for all uses with attribution encouraged.
This project is in early development, but the foundations are strong. If the ideas resonate, you’re invited to contribute, refine, or build your own implementation.
Check out:
/docs/traps-and-challenges.md— common failure points and how to address them/docs/example-akus.md— starter templates for AKU formulation/docs/agent-blueprint.md— example agent design using AKU-RIM principles
Or start a conversation in Issues to open a new trail.