The Inference Report

The dominant theme across today's repos is agents with memory and structured knowledge. volcengine/OpenViking and akitaonrails/ai-memory both tackle the same problem: agents forget. OpenViking frames it as a unified context database that handles memory, RAG, and skills together. ai-memory approaches it differently, solving for persistence across agent CLI sessions and vendor handoffs. Both recognize that stateless LLM calls don't work for real workflows. The cybersecurity skills repo (Anthropic-Cybersecurity-Skills) is the most concrete instantiation of this pattern: 817 structured skills mapped to six frameworks, designed to work across twenty platforms. It's not trending because of hype. It's trending because teams building agents need a way to give them domain knowledge that actually sticks.

Alongside memory, there's a secondary wave of agentic frameworks trying to codify methodology itself. mattpocock/skills and obra/superpowers both position themselves as skill libraries for developers, though superpowers frames it explicitly as "software development methodology." These are less about infrastructure and more about capturing what good agents should know how to do. That matters because it suggests the industry is past the "agents are just prompts" phase and into the "agents need structured knowledge and repeatable patterns" phase. The trading engine (nautilus_trader) and self-hosted photo management (immich) show the pattern extends beyond AI: production systems need deterministic behavior, efficient memory use, and local-first operation. On the inference side, both omlx and mlx-serve solve the same problem for Apple Silicon specifically, suggesting that on-device LLM serving is becoming table stakes for agent platforms. The discovery repos reinforce this: Talkify adds voice input as a primitive, remove-ai-watermarks handles provenance metadata, and Trilium offers personal knowledge bases. Each solves a piece of the agent stack that wasn't there six months ago.

Jack Ridley