ActiveGraph: continuity layer for long-running AI agents

Who: Posted by @yohei (AI agent researcher and builder), writing in his own voice as both poster and author of the linked article.

What's new: Yohei proposes ActiveGraph, an architectural pattern for giving long-running AI agents a persistent memory and continuity layer. The core idea is that an agent's entire working state — history, goals, partial results — lives in a structured knowledge graph that survives across sessions, rather than being crammed into a single prompt window that gets wiped when the session ends.

How it works: The design draws on older computing patterns — Yohei explicitly compares it to architectures predating modern AI — to treat the agent's ongoing task as a live, queryable graph of nodes and relationships rather than a linear conversation log. Instead of the agent re-reading a long transcript to figure out where it left off, it queries the graph for its current state, pending steps, and relevant past decisions. This is analogous to a contractor keeping a project board pinned to the wall rather than re-reading every email thread each morning.

Caveats: The article is a concept piece; Yohei notes code is forthcoming. There are no benchmarks, no comparisons to existing agent memory frameworks, and no ablation data showing the graph approach outperforms simpler methods like summarization or retrieval-augmented memory. The claim of novelty is also self-qualified — "I might be going crazy" — so the idea's true originality relative to prior work in knowledge graph-backed agents remains unconfirmed.