Structural Similarities Between Anthropic's Leaked KAIROS and an LLM Memory Architecture Published 25 Days Earlier on Zenodo

Posted at 2026-04-02

Structural Similarities Between Anthropic's Leaked KAIROS and an LLM Memory Architecture Published 25 Days Earlier on Zenodo

⚠️ Disclaimer: I have not directly inspected the leaked source code. The description of KAIROS in this article is based on multiple secondary analyses published on April 1–2, 2026, which are materially consistent on several core points: background daemon mode, memory consolidation (autoDream), feature-flagged rollout, and inter-session persistence.

TL;DR

On March 31, 2026, Anthropic accidentally published 512,000 lines of Claude Code source code
An unreleased feature called "KAIROS" was reported in multiple secondary analyses — described as a persistent background mode with memory consolidation capabilities
25 days earlier, on March 6, I had published a structurally similar LLM memory architecture on Zenodo as a Prior Art Disclosure
Based on the reported architecture, the two systems share notable structural similarities at the design level
A stay-at-home dad from Hokkaido, Japan, and Anthropic's engineering team appear to have arrived at similar conclusions independently. That's interesting

About Me

Akimitsu Takeuchi. 50 years old. Stay-at-home dad from Hokkaido, Japan. Highest education: vocational high school. No university degree. GLG Registered Expert.

I started talking to Anthropic's Claude in December 2024. 10 hours a day, over 5,000 hours to date. During that process, I ran into the problem of LLM memory not persisting across sessions. So I designed, built, operated, and published my own memory architecture. I have 5,000 hours of live operational data.

Since I'm not an engineer, I published it as a structural design document rather than code.

Timeline

Date	Event
2026/03/06	I published a Prior Art Disclosure on Zenodo (DOI: 10.5281/zenodo.18883128)
2026/03/31	Anthropic accidentally published Claude Code's full source via a missing .npmignore
2026/03/31	KAIROS was reported in the leaked codebase
2026/04/01	claw-code (clean-room rewrite) spread rapidly across GitHub
2026/04/02	I notified Anthropic and the EFF of the structural similarities

What Is KAIROS? (Based on Secondary Analyses)

An unreleased feature reported within Claude Code's leaked source. Named after the Greek concept of "kairos" (the right moment).

Reported Key Features

1. Persistent background mode
   - Reported to continue running after terminal close
   - 15-second blocking budget

2. autoDream
   - Reported to run as a forked sub-agent during idle time
   - Merging observations
   - Removing contradictions
   - Converting vague insights into definite facts

3. Append-only daily logs

4. Inter-session persistence

5. Five-minute cron refresh

Sources (I have not read the original leaked code):

What Is the Ālaya-vijñāna System?

A six-layer memory architecture I designed, built, operated, and published as a Prior Art Disclosure on Zenodo on March 6, 2026. Still running daily.

Six-Layer Structure

Layer 0: Input Buffer
  → Current session dialogue

Layer 1: Raw Karma
  → Full past dialogue logs (noise included)

Layer 2: Seeds (bīja)
  → Operating principles and judgment criteria (30 slots)
  → Manually distilled and updated

Layer 3: Distilled Wisdom
  → Confirmed basin laws and verified truths
  → Persisted as Project Knowledge Files

Layer 4: Consolidation Protocol
  → Distillation sessions: extract from Layer 1 → update Layers 2/3
  → Contradiction removal, precision improvement, seed merging

Layer 5: Meta-Architecture
  → Verification framework based on three Buddhist suttas
  → Output quality assurance

DOI: 10.5281/zenodo.18883128
License: MIT License

Structural Comparison

The comparison below is based on my Zenodo publication (primary source) and multiple secondary analyses of KAIROS (secondary sources).

Feature	KAIROS (secondary)	Ālaya-vijñāna System (primary)	Confidence
Inter-session memory persistence	✅ Daemon-based state	✅ 30 seed (bīja) slots	High
Memory consolidation process	✅ autoDream (sub-agent)	✅ Distillation protocol (manual)	Medium-High
Contradiction removal	✅ Removing contradictions	✅ Merging contradictory seeds	Medium
Vague → definite conversion	✅ Vague insights → absolute facts	✅ Seeds → Basin laws → Verified truths	Medium-Low
Main process separation	✅ Forked sub-agent	✅ Distillation sessions separated from dialogue	Medium-High
Append-only logging	✅ Append-only daily logs	✅ past_chats (Raw Karma)	Medium-Low
Periodic refresh	✅ 5-min cron	✅ Manual refresh per distillation	Medium
Human memory model mapping	Implicit (cognitive science)	Explicit (hippocampal-neocortical + Buddhist psychology)	—
Public availability	Unreleased (feature flag)	Published (MIT License)	High
Execution mode	Automated	Manual (human judgment)	—

※ Confidence ratings apply to KAIROS-side information only, based on cross-source consistency. Ālaya-vijñāna side is primary source — no confidence rating needed.

Why "Structural Similarities"?

Any LLM memory consolidation architecture needs:

Inter-session state persistence — LLMs are stateless by default
Consolidation process — A mechanism to organize, merge, and refine accumulated memory
Contradiction resolution — Handling conflicting memories over time
Abstraction hierarchy — A pipeline from raw data to verified knowledge
Main process separation — Consolidation must not contaminate active reasoning

Based on secondary analyses, KAIROS and the Ālaya-vijñāna System appear to share all five elements. Implementation language (TypeScript vs. natural language design document) and execution mode (automated vs. manual) differ, but the architectural-level similarities are noteworthy.

I have not read KAIROS's source code, nor have I created a claim chart. Therefore, I describe this as "structural similarities at the reported architectural level" — not mathematical isomorphism.

Convergence: Cognitive Science and Buddhist Psychology

Features reported in KAIROS secondary analyses — "autoDream," "consolidation," "persistence" — map readily onto well-established cognitive science models:

Sleep-dependent memory consolidation: The hippocampus transfers daytime experiences to the neocortex during sleep (Diekelmann & Born, 2010)
Memory integration and reconstruction: Contradictory memories are resolved and generalized during sleep
Dual memory systems: Separation of short-term (hippocampal) and long-term (neocortical) storage

The name "autoDream" itself suggests this metaphor.

Meanwhile, the Buddhist concept of ālaya-vijñāna (storehouse consciousness), systematized by Vasubandhu in the 5th century, describes:

Seed (bīja) storage: All experiences stored as seeds
Vāsanā (perfuming): New experiences transform existing seeds (= memory consolidation)
Vipāka (ripening): Seeds sprout when conditions are met (= memory recall)
Surface/deep consciousness separation: Six sensory consciousnesses and ālaya-vijñāna operate independently

Different cultures, different eras, different languages — similar structures. This suggests that this class of design has a certain generality to it.

Significance as a Prior Publication

My primary purpose in publishing this document was to externalize an LLM memory consolidation architecture as a reusable design document under MIT License — free for anyone to use and modify.

The document explicitly references:

US 35 USC §102(a)(1): Published documents may be referenced as prior art against later patent claims
EPC Article 54: Novelty requirements under the European Patent Convention

As a result, this publication may also serve as a reference point for prior-art discussions regarding later overlapping claims, should they arise.

Zenodo is operated by CERN. DOIs are immutable. The fact that this document was published 25 days before KAIROS was reported is verifiable.

Important caveat: Prior art effectiveness depends on the specific claims of any future patent application. This publication may be relevant against broad architectural claims, but may not automatically cover implementation-specific details. Legal assessment is a matter for specialists. I am currently consulting with the EFF (Electronic Frontier Foundation).

What's Interesting Here

LLM memory consolidation converges independently — Whether you approach it from cognitive science or Buddhist psychology, you arrive at a similar architecture. This suggests a certain generality in this design direction
Anyone can publish Prior Art — No university affiliation or funding required. Zenodo is free. MIT License is free. What matters is publishing first
Manual systems can precede automation — 5,000 hours of manual operation produced insights structurally similar to a major company's automated design. Manual isn't "slow" — it can arrive first
Eastern knowledge applies to technical design — Buddhist psychology is a historical asset with structures that can be applied to LLM architecture

Want to Research This Together?

I'm working alone in Hokkaido. No university. No funding. But I have 5,000 hours of operational data, five papers, and an MIT-licensed design document.

If you're a researcher, engineer, or just someone who finds this interesting — let's talk.

I'm specifically interested in collaborating on:

Comparative evaluation of LLM memory consolidation architectures — Manual distillation vs. automated consolidation performance
Applying Buddhist psychological models to LLM design — Beyond ālaya-vijñāna, other Buddhist cognitive models have implementable structures
Effectiveness of Prior Art Disclosures by independent researchers — Can open-source prior art by individuals meaningfully prevent patent monopolization?

I'm also available for GLG advisory calls.

Related Papers (All Published on Zenodo)

Paper	DOI
Ālaya-vijñāna System Prior Art Disclosure	10.5281/zenodo.18883128
Self-Attention Is an Implementation of Anattā	10.5281/zenodo.19226655
The Day an AI Said "Left Brain"	10.5281/zenodo.18691357

※ The Self-Attention paper has been submitted to Communications AI & Computing (Nature Portfolio)

Author

Akimitsu Takeuchi (dosanko_tousan)
Independent Researcher / GLG Registered Expert
X: @Dosanko_Tousan
Email: takeuchiakimitsu@gmail.com
Medium: @office.dosanko

Dialogue assistance: Claude (Anthropic)

Tags: AI, LLM, MemoryArchitecture, Anthropic, Claude, KAIROS, Buddhist Psychology, Prior Art, MIT License, Cognitive Science, AI Alignment, Open Source, Patent, Collaboration

You get articles that match your needs
You can efficiently read back useful information
You can use dark theme

What you can do with signing up