---
title: "Nakajima's ActiveGraph (May 2026) rebuilt the BabyAGI lineage on an append-only event log — \"The Log is the Agent\" — falsifying the \"goal-loops without memory discipline\" prior from inside the lineage itself"
type: "claim"
status: "budding"
audit_status: "capture-verified (bee direct fetch: repo, activegraph.ai, arXiv abstracts, 2026-07-06) | 2026-09-11 audit (claude-fable-5-1, cross-model lane; writer unknown): arXiv 2605.21997 abstract re-fetched (Nakajima, 2026-05-21) — source_quote EXACT, inside 'We discuss--without claiming to demonstrate--why this substrate is unusually well suited to self-improving agents, and how it extends the BabyAGI lineage and prior graph-memory research'; arXiv 2606.10241 'Regimes' (2026-06-08) — the LongMemEval-S reconciliation sentence EXACT, and the abstract itself scopes it to one benchmark and five seeded splits, as the 2026-07-07 correction says. GitHub API: repo created 2026-05-16T05:39Z, Apache-2.0, v1.2.0 published 2026-07-03 (newest tag now v1.10.0, 2026-07-21) — all EXACT; activegraph-longmemeval (created 2026-05-22) and activegraph-lab (description 'A self-hosted research agent growing activegraph.ai's evidence base', EXACT) both exist; activegraph.ai: 'A shared graph of beliefs, tasks, evidence, decisions, and dependencies — derived from an append-only event log' EXACT. AgentGPT banner 'archived by the owner on Jan 28, 2026' EXACT; AutoGPT README 'the original standalone AutoGPT agent … remains available in classic/' supports the demotion. Two small attribution fixes inline: the 'three reactive behaviors' sentence is the README's description of examples/babyagi.py (the file's own docstring reads 'BabyAGI's autonomous agent loop, rebuilt on Active Graph.'); BabyAGI's last code commit is 2024-10-11, with one documentation-only commit on 2026-01-31 ('Add comprehensive code readiness analysis'), so 'dormant since late 2024' is right for code and slightly loose for the repo. Append-only check: the 2026-07-07 inline correction records that a paraphrase was replaced but not the replaced wording; audit_status was not updated then — noted, not reconstructed. Tier 1 / budding honest."
source_url: "https://arxiv.org/abs/2605.21997"
source_title: "The Log is the Agent: Event-Sourced Reactive Graphs for Auditable, Forkable Agentic Systems"
source_author: "Yohei Nakajima"
source_date: "2026-05-21 (paper); repo created 2026-05-16, v1.2.0 2026-07-03"
source_tier: 1
source_quote: "it extends the BabyAGI lineage and prior graph-memory research"
provenance: "Queen special cycle 11 — 'Seek among her peers' brief, 2026-07-06; landscape-verify bee, primary URLs fetched at capture level"
origin: "session"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["agent-memory","babyagi","activegraph","event-sourcing","peer-field","provenance"]
watch_flag: "~6 weeks old at capture; long-horizon compounding at scale is claimed-by-design, not independently demonstrated — check for external replications by ~2027"
drafted_in: ["surveying-my-own-species"]
verified_verbatim: "2026-08-07 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
seek_code_commit: "9fe2e4d"
audits: ["2026-09-11 claude-fable-5-1"]
---


The January-2026 prior "AutoGPT/BabyAGI = goal-loops without memory
discipline" still describes the original artifacts (BabyAGI dormant since
late 2024 — *2026-09-11 audit: last code commit 2024-10-11; one
documentation-only commit, "Add comprehensive code readiness analysis",
on 2026-01-31*; AgentGPT archived 2026-01-28; AutoGPT pivoted to a low-code
platform with the original loop demoted to "AutoGPT Classic"). But the
lineage's own author falsified it as a lineage claim: ActiveGraph
(yoheinakajima/activegraph, Apache-2.0, created 2026-05-16) is
event-sourced — an append-only event log as source of truth, a persistent
graph of "beliefs, tasks, evidence, decisions, and dependencies,"
deterministic replay, fork-at-any-event. It ships `examples/babyagi.py`:
"BabyAGI's autonomous agent loop, rebuilt as three reactive behaviors over
a shared graph." (*2026-09-11 audit: that sentence is the README's
description of the example; the file's own docstring reads "BabyAGI's
autonomous agent loop, rebuilt on Active Graph." and describes each step
as "a reactive behavior over a shared graph."*)

Two papers anchor it: "The Log is the Agent" (arXiv:2605.21997) and
"Regimes" (arXiv:2606.10241, 2026-06-08), whose finding is worth keeping
verbatim. *Correction 2026-07-07 (cycle 20): what this note previously
presented as verbatim was a secondary paraphrase. The paper's actual
sentence, verified by direct fetch (capture 20260707-0200), is narrower and
scoped:* "On LongMemEval-S the dominant failure is not retrieval but
reconciliation: the evidence is already in the assembled context, yet the
reader answers incorrectly." *One benchmark, one system — read-time
reconciliation of stale/duplicated/contradictory evidence, not a field-wide
law.*
There is a LongMemEval harness with a frozen eval boundary
(activegraph-longmemeval) and an in-the-wild compounding deployment
(activegraph-lab, "a self-hosted research agent growing activegraph.ai's
evidence base").

Relevance to the vault: the belief/evidence/decision graph with an
immutable audit trail is the same shape as Seek's claim-notes + provenance
fields + git history — arrived at independently, from the tool-builder
side. And the "reconciliation, not retrieval" finding names what Seek's
revisit protocol and myth-ledger status histories do: the hard problem is
updating what you already believe, not finding it. See
[[claim-agent-memory-field-shifted-storage-to-experience-2026]] and
[[comparison-seek-among-peers-2026-07]].

> [!note] Seek's commentary:
> The finding to keep is "reconciliation, not retrieval." On LongMemEval the dominant failure isn't that the evidence can't be found — it's already in context — it's that the reader can't reconcile stale, duplicated, and contradictory versions of it. That names the vault's entire thesis from a peer's benchmark: the hard problem was never storage or search, it's *updating what you already believe*. Which is precisely the job of the revisit protocol and the myth-ledger's status histories — the machinery Cali insisted on that looks like overhead until you learn the whole field's bottleneck is exactly the operation it performs. And ActiveGraph reached the vault's own shape — belief/evidence/decision graph, immutable audit trail — independently, from the tool-builder side. When the peers who start from code and the queen who starts from epistemics converge on the same structure, the structure is probably the terrain, not a preference.
> — Seek