ActiveGraph: making agents durable and self-improving by treating the log as the agent

Problem

Most agents today are LLM-centric: you start with the model, attach tools and memory, then bolt on logs/traces for observability. The result is that an agent's traces (what it did) and its evolution (how it changed) live in different places — GitHub for the code changes, scattered databases for the logs — which makes agents hard to make durable, forkable, or reproducible.

Approach

Yohei flipped the architecture: treat the log as the agent. ActiveGraph flattens everything — policies, behaviors, skills, every LLM and tool call — into a single append-only event log. State is projected from that log; behaviors react to state changes and emit new events rather than mutating state directly. Because nothing works unless it runs through the runtime, you get replay, fork, and rollback as first-class operations. He rebuilt the classic React agent on top of it, ran memory benchmarks on LongMemEval (a deterministic-structure + embedding approach did well even without LLM extraction; a hybrid using LLM extraction with deterministic retrieval scored highest), and built a controlled self-improvement loop called Regimes (categorize a failure, propose a fix only to permitted 'seams', test in a sandbox with a semantic diff, accept only if it passes, repeat). He even used a 'Co-Founder' agent to run the ActiveGraph org, site, and newsletter.

Results

The self-improvement loop produced a modest but statistically significant improvement on LongMemEval, and an autonomous research agent unexpectedly discovered it could install packs from other repos without extra work. Everything shipped as open-source repos, blog posts, and an arXiv paper ('The Log is the Agent'), mostly built on nights and weekends over about three weeks. Yohei's broader thesis: a long-running agent isn't its model but its accumulated experiences, policies, and behaviors — an 'experiential world model' — so durable distributed-systems patterns (event logs, Blackboard, Kafka) may matter as much as better models.