The future of AI is not scale.
Status: Proposed
Date: 2026-02-18
Author: System Architect (AgentDB v3)
Supersedes: None
Related: ADR-003 (RVF Format), ADR-006 (Unified Self-Learning RVF), ADR-007 (Full Capability Integration), ADR-008 (Chat UI RVF)
Package: @agentdb/causal-atlas
RVF — RuVector Format
One file. Store vectors. Ship models. Boot services. Prove everything.
Status: Proposed Date: 2026-02-10 Authors: RuvNet, Claude Flow Team Version: 1.0.0 Related: ADR-006 (Unified Memory), ADR-009 (Hybrid Memory Backend), ADR-027 (RuVector PostgreSQL), ADR-048 (Auto Memory Integration), ADR-049 (Self-Learning Memory GNN)
ruvnet
/ claude-flow-vs-claude-code.md
Last active
March 4, 2026 22:46
Architectural Comparison: Claude Flow V3 vs Claude Code TeammateTool
Date: 2026-01-25 Analysis: Side-by-side comparison of Claude Flow V3 swarm architecture (developed by rUv) and Claude Code's TeammateTool (discovered in v2.1.19)
A detailed analysis reveals striking architectural similarities between Claude Flow V3's swarm system and Claude Code's TeammateTool. The terminology differs, but the core concepts, data structures, and workflows are nearly identical.
ruvnet
/ coherence_engine.md
Last active
March 1, 2026 16:12
A ruvector-based coherence engine that models state as a constrained graph, detects structural inconsistency via sheaf Laplacian residuals, and gates actions with auditable, deterministic witnesses instead of probabilistic confidence.
Most systems try to get smarter by making better guesses. I am taking a different route. I want systems that stay stable under uncertainty by proving when the world still fits together and when it does not. That is the point of using a single underlying coherence object inside ruvector. Once the math is fixed, everything else becomes interpretation. Nodes can be facts, trades, vitals, devices, hypotheses, or internal beliefs. Edges become constraints: citations, causality, physiology, policy, physics. The same residual becomes contradiction energy, and the same gate becomes a refusal mechanism with a witness.
This creates a clean spectrum of applications without rewriting the core. Today it ships as anti hallucination guards for agents, market regime change throttles, and audit ready compliance proofs. Next it becomes safety first autonomy for drones, medical monitoring that escalates only on sustained disagreement, and zero trust security that detects structural incohe
I’ve put together a new tutorial for RV Lite and RuVector that reflects how I actually work. Prediction by itself is noise. Knowing what might happen is useless if you cannot adapt, respond, and steer toward the outcome you want.
This system is about doing all three. It does not stop at forecasting a future state. It models pressure, uncertainty, and momentum, then plots a viable course forward and keeps adjusting that course as reality pushes back. Signals change, competitors move, assumptions break. The system notices, recalibrates, and guides the next step.
What makes this different is where and how it runs. RV Lite and RuVector operate directly in the browser using WebAssembly. That means fast feedback, privacy by default, and continuous learning without shipping your strategy to a server. Attention mechanisms surface what matters now. Graph and GNN structures capture how competitors influence each other. Simulations
This guide explains four powerful mathematical techniques that will differentiate ruvector from every other vector database on the market. Each solves a real problem that current databases can’t handle well.
Designing a Rust-Based Long-Term Memory System (MIRAS + RuVector)
Building a long-term memory system in Rust that integrates Google’s MIRAS framework (Memory as an Optimization Problem) with the principles of RuVector requires combining theoretical insights with practical, high-performance components. The goal is a memory module that learns and updates at inference-time, storing important information (“surprises”) while pruning the rest, much like Google’s Titans architecture  . We outline a modular design with core components for surprise-gated memory writes, retention/forgetting policies, associative memory updates, fast vector similarity search, and continuous embedding updates. We also suggest Rust crates (e.g. RuVector) that align with geometric memory, structured coherence, and update-on-inference principles.
Memory Write Gate (Surprise-Triggered Updates)
A surprise-based write gate decides when new information merits permanent storage. In Titans (which implements MIRAS), a “surprise metric” measur
ruvnet
/ PowerInfer.txt
Created
January 5, 2026 02:09
PowerInfer-Style Activation Locality Inference Engine for Ruvector (SPARC Specification)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| PowerInfer-Style Activation Locality Inference Engine for Ruvector (SPARC Specification) | |
| Specification | |
| Goals and Motivation: The goal is to create a high-speed inference engine that exploits the activation locality in neural networks (especially transformers) to accelerate on-device inference while preserving accuracy. Modern large models exhibit a power-law distribution of neuron activations – a small subset of “hot” neurons are consistently high-activation across inputs, while the majority are “cold” and only occasionally activate . By focusing compute on the most active neurons and skipping or offloading the rest, we can dramatically reduce effective model size and latency. The engine will leverage this insight (as in PowerInfer ) to meet edge deployment constraints. Key performance targets include multi-fold speedups (2×–10×) over dense inference and significant memory savings (e.g. 40%+ lower RAM usage ) with minimal accuracy impact (<1% drop on benchmarks ). It should enable running larger models |
NewerOlder