How Synthetic Players Used ActiveGraph to Verify 4,919 Runs Without New Model Calls
A technical account of recording language-model experiments as events, replaying them from stored responses, and preserving failures and provenance for review.
> activegraph / blog
Notes from the active graph.
Announcements, deep-dives, and research from the team building activegraph. RSS at /blog/rss.xml.
Log · 19 entriesNewest first
Pokemon TCG AI Battle - Things we tried that didn't workexternal
I played the card game over 20 years ago, and these days play with AI a lot. I took a stab at this tournament without relearning the game to see how far AI could take me. The short version: some things I tried improved weak deck/agent scores a bit, but the best scores came from copying strong public examples from the forum, and I couldn't meaningfully improve them. The write-up documents what moved, what failed, and what I would not spend games on again — the constraints (5 submissions/day, ~24hr score settling, incomplete offline opponents, stochastic outcomes, Kaggle runtime drift) shape most of the lessons.
ActiveGraph 1.2.0 to 1.7.1: hardening the reference packs pulled the runtime forward
A coordinated push across two repos: the reference packs got hardened for autonomous-agent use, and the activegraph runtime shipped seven releases in a day to support them. The through-line is one loop — fork, trial in a child process, owner approval against pinned bytes, quiescent promote, watch window — recorded end-to-end in the event log.
Five Numbers and a Negative Resultexternal
Part 1 of a series on benchmarking activegraph as the memory substrate for a terminal agent. activegraph works as a harness (competitive with terminus control), but naive graph retrieval lost to the prompt-cached transcript baseline. The interesting signal is in the shape of the failures.
Active Graph v1.2.0: The Graph Projection Becomes Pluggable
v1.2.0 makes the materialized graph a pluggable seam with FalkorDB as the first external backend — native edges, Cypher query push-down, a conformance suite for future backends — plus a CI gate for the full test suite and one license everywhere.
ActiveGraph Packs: A Composable Assistant Substrate
A reference pack library for graph-native assistants — memory, identity, tools, comms, and domain capabilities composing through evented state instead of a central orchestrator.
I Built a Research Agent That Shows Its Work
A graph-native, event-driven deep research agent that exposes its reasoning, evidence, and dead ends instead of flattening them into prose. Open source, live demo.
Building Deep Research on ActiveGraph: A Field Report
A first-person field report on building ActiveGraph Deep Research — the mental shifts, the bugs that taught me the runtime, and the design decisions I'd defend.
Active Graph v1.1.0: Better Runtime Resilience, Better Operator Tools
v1.1.0 tightens the surfaces that matter once an agentic system leaves the demo stage: bounded LLM retries with audit, CLI inspect tools, fork-time pack overrides, OpenTelemetry metrics, OpenAI tool-call parity, and stricter docs and release gates.
Regimes: An Auditable, Held-Out Gated Improvement Loop Demonstrated on LongMemEval with ActiveGraphexternal
Paper #2 follows up on paper #1's open question — ActiveGraph as an architectural affordance for self-improving agents. A reproducible gated self-improvement loop with modest autonomous gains on LongMemEval, plus a regime-to-seam approach where categorized failures fix targeted areas behind a binding held-out gate. Discussion: https://x.com/yoheinakajima/status/2064830891143381076
AG Coder: A Reference Coding Agent Built on ActiveGraph
An open-source coding agent built on ActiveGraph: graph-native, event-sourced, and auditable by default. A reference for what a coding agent looks like when the runtime is the substrate, not a wrapper around an LLM loop.
Introducing learn.activegraph.ai
A new site for activegraph that teaches the framework as a refactor: start from the while-loop agent everyone has built, substitute one piece at a time, and end up at an event-sourced graph runtime with replay, provenance, contradictions, and forkable conclusions.
Code Without Authority
An authority lifecycle for generated behavior in an event-sourced agent runtime
Compile From the Log, Don't Replace It
Hybrid Semantic Memory on LongMemEval-S: Retrieval Significance, Reader Bottlenecks, and the Case for Provenance-Backed Facts
Self-Modification Is Not Binary
Typed, auditable self-change on an event-sourced agent runtime
What Semantic Memory Forgets: ActiveGraph on LongMemEval-S
Debugging ActiveGraph on LongMemEval-S: LLM extraction moved the semantic-memory pack from 60.6% to 83.4%, role-aware assistant retention fixed a 75.0% → 98.2% assistant-memory failure class, and the remaining errors showed why semantic memory should compile back to source evidence.
Evidence Compilation Before Semantic Memory: ActiveGraph on LongMemEval-S
Technical benchmark note: 85.6% QA accuracy and 86.2% turn answer-in-context at 2,462 mean context tokens, with deterministic non-generative ingestion.
The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systemsexternal
Append-only event log as source of truth, working graph as deterministic projection, behaviors react and emit, yields deterministic replay + cheap forking + end-to-end lineage. Discussion: https://x.com/yoheinakajima/status/2057812713045377055
activegraph: agents that coordinate through replayable stateexternal
v1.0 is out: a Python runtime where long-running agents share a reactive, event-sourced graph as their world — fork, replay, and audit included.