Yohei Nakajima (author of BabyAGI) has published a paper on ActiveGraph, an architecture for agents that starts from the log rather than the language model. In most frameworks, the agent is a conversational loop with tools, rules, and a log added on top for observability. Here the append-only log is the source of truth, the work graph reconstructs deterministically from the log, and behaviors (functions, classes, LLM routines) react to graph changes by emitting new events. No component gives orders to others: everything flows through the shared graph.

This inversion gives you three things that retrieval-based memory systems don’t: deterministic replay of any execution from its log, cheap forks that branch an execution at any point without re-running the shared prefix, and end-to-end lineage from a high-level goal down to the single model call that produced each artifact. If you’re building agents that need to correct themselves, debug past decisions, or start from intermediate points, this structure lets you do it without rewriting your infrastructure each time.

The paper includes a due diligence example whose complete causal structure reconstructs only from the log. The implementation is open source (Apache 2.0) with reproducible demo, deterministic replay, fork-and-diff, and lineage tracking. If you work with complex agents or want to understand how the tools you use work, it’s worth looking at the architecture: it changes what you can inspect and what you can reuse.