Mem-on-the-Go

Agents forget. Every session starts cold, every debugging run re-discovers the same patterns, every code review loses context the moment the process exits. Mem-on-the-Go fixes that. A universal memory system that travels with the agent — browser, Node.js, edge function, fully offline — with zero setup and no API keys.

Built on RuVector, the system automatically selects the highest-performance backend available. Native Rust with SIMD acceleration delivers 61 microsecond search latency and 50K+ QPS when available. Drop down to a 5.5KB WASM runtime in the browser and it still works, entirely client-side, with the bundled all-MiniLM-L6-v2 model generating embeddings without touching a server.

Nine reinforcement learning algorithms run continuously: Q-Learning, SARSA, DQN, PPO, Actor-Critic, Policy Gradient, Decision Transformer, MCTS, and Model-Based RL. They train together through contrastive learning with InfoNCE loss and hard negative mining, improving retrieval quality 36% over time. LoRA adapters handle fine-tuning efficiently. EWC++ prevents catastrophic forgetting of patterns the system already learned.

Hybrid search combines BM25 keyword matching with HNSW vector similarity through Reciprocal Rank Fusion — exact term matches and semantic understanding working together. A GNN with 8-head multi-head attention reranks candidates for a 12.4% recall improvement. SONA auto-selects from 7 attention strategies per query in under 100 microseconds.

The .rvf cognitive container format bundles index, embeddings, learned RL state, and graph structure into a single portable file. Agent trajectory tracking captures real behavior — debugging sessions, code reviews, research workflows — and distills it into retrieval patterns the system actually improves from. The agent learns from itself.