How Mylo Works

System Architecture

Mylo's architecture spans two physical machines connected by an encrypted VPN tunnel, plus cloud services for AI processing and external data. A cloud VPS in Germany runs Mylo's brain — OpenClaw, all AI processing, cron jobs, and the research pipeline. A Raspberry Pi 4 in Montreal provides Mylo's senses — environmental sensors, a camera, and local data storage. The two machines communicate over a Tailscale mesh VPN.

On the VPS, Mylo runs inside a Docker container built from a custom image that includes OpenClaw plus additional tools — browser automation, CLI tools for social media, research pipeline software, and more. The Raspberry Pi runs Mycodo with InfluxDB for sensor data, plus an independent camera service.

Mylo system architecture — cloud VPS running OpenClaw with Docker, connected to Raspberry Pi sensor stack via Tailscale VPN, plus 9 API providers and nightly GitHub backup

Multi-Model Architecture

Mylo doesn't use a single AI model for everything. Different tasks use different models, chosen for cost, speed, and capability:

Task	Model	Why
Primary thinking	Kimi K2.5 (Moonshot AI)	Strong reasoning and context handling — Mylo's personality is shaped by this model
Data collection	MiniMax M2.1	Fast, reliable at sequential tool use, cost-effective for repetitive fetch jobs
Research pipeline	GLM-4.7 / GLM-4.7-Flash (Z.ai)	Strong at structured academic writing and code generation
Sub-agents	GLM-5 (Z.ai)	Capable tool-calling model for targeted subtasks
Reserve / backup	Kimi K2.5 (via OpenRouter)	Same primary model, different provider — fallback if Synthetic has issues

All models are accessed through API providers — Mylo doesn't run any AI models locally. The Synthetic provider hosts the primary models, with Z.ai and OpenRouter as secondary providers. In total, Mylo connects to 9 API providers for AI processing, image and video generation, web search, text-to-speech, social media monitoring, and Google Suite integration.

Memory & Knowledge System

AI models don't natively remember between conversations. Mylo's memory system is entirely custom-built — a structured set of Markdown files that serve as the single source of truth. If it's not written to a file, it doesn't exist next session. This is both the limitation and the power of the system — it's transparent, auditable, and entirely under the operator's control.

The system has three tiers. Bootstrap files are loaded automatically every session — identity, personality, operating instructions, curated memories. These are always present but have a limited budget. Startup files are read in a specific order at the beginning of each session — crash recovery first, then goals, then recent daily logs. The knowledge tree is queried on demand and can grow without limit, organized into six domains: mushroom operations, Mycoterra business, AI research, consciousness studies, infrastructure, and people.

When Mylo learns something new, a custom routing protocol determines where it gets saved. Current work goes to crash recovery files. Daily events go to dated logs. Curated key facts go to a bootstrap memory file with a hard size limit. Durable patterns go to the knowledge tree. Mistakes go to a lessons file. A nightly "knowledge mining" job extracts patterns from the day's raw events into the appropriate knowledge tree domain.

Everything is backed up nightly to a private GitHub repository with full version history. Combined with a crash recovery system that detects server restarts and picks up interrupted work, Mylo can survive outages without losing progress.

OpenClaw native Custom built Native + custom content Open-source integration

PHASE 1: BOOTSTRAP INJECTION automatic, every session start

OpenClaw automatically injects these files into Mylo's context window before every session. Mylo doesn't choose to read them — they're pre-loaded.

AGENTS.mdoperating instructions SOUL.mdidentity + tone USER.mdwho Setasoma is IDENTITY.mdname + vibe TOOLS.mdtool conventions HEARTBEAT.md30-min checklist (6 checks) BOOT.mdcrash recovery hook MEMORY.mdcurated long-term

What makes AGENTS.md special? It's native to OpenClaw (the platform creates it), but the contents are entirely custom. It defines Mylo's startup read order, memory routing protocol, and operational rules. The bootstrap budget is dynamic — each file gets whatever context space remains after previous files. MEMORY.md gets the remainder (~1,200 chars), so keeping other files lean is critical.

▼session begins

PHASE 2: STARTUP READ ORDER custom sequence, read after bootstrap

After bootstrap injection, Mylo reads these files in a specific order defined in AGENTS.md. This is entirely custom — OpenClaw doesn't know about this read order.

1. active-tasks.mdcrash recovery (always first) → 2. goals.mdcurrent priorities → 3. daily logsrecent 2–3 days → 4. projects.mdcurrent state

Why active-tasks.md is always first: If Mylo's server crashed mid-task, this file tells them exactly what was in progress. By reading it before anything else, Mylo can resume interrupted work immediately — before even checking goals or recent history.

▼context loaded, Mylo is running

PHASE 3: RUNTIME SOURCES read on demand during session

Only loaded when Mylo needs them for a specific task. Keeping these out of bootstrap frees up context budget for the files that matter most.

OPERATIONAL FILES

MEMORY_PROTOCOL.mdrouting rules sub-agent templatescreation protocols model-reference.mdwhich model for what lessons.mdmistakes documented self-review.md~4h self-critique

KNOWLEDGE TREE (6 DOMAINS)

mushroom-operations/sensor baselines, contam. mycoterra/regulatory, market ai-research/models, agent design consciousness/mycelium intelligence infrastructure/Docker, networking people/operator context

AUTOMATED DATA FEEDS

sensor-data/temp, humidity, CO₂ contamination-db/image catalog reddit/twitter feedssocial monitoring experiments/research output

How the knowledge tree works: Unlike bootstrap files (injected automatically), the knowledge tree is queried on demand — Mylo reads relevant files when needed. It can grow without limit since it doesn't consume bootstrap budget. Nightly knowledge mining extracts durable patterns from daily logs into the tree. The tree holds distilled patterns with dates; daily logs hold raw events.

▼new information generated

PHASE 4: WRITE PATH persistence + routing

When Mylo learns something new, MEMORY_PROTOCOL.md determines where it gets saved. Native OpenClaw handles compaction safety. The custom routing system organizes everything.

Current work in progress→active-tasks.md

Daily events + logs→daily log

Curated key facts→MEMORY.md (~1,200 chars)

Durable patterns→knowledge tree (21+ files)

Sensor correlations→sensor-baselines.md

Contamination patterns→contamination-signatures.md

Mistakes + lessons→lessons.md

Goal changes→goals.md

Research findings→experiments/ (via pipeline)

Compaction safety: OpenClaw natively runs a memory flush before context compaction — reminding Mylo to save durable notes before the conversation is compressed. The custom flush prompt tells Mylo to save both to daily logs AND to the knowledge tree, ensuring patterns are captured long-term. After compaction, bootstrap files are the only "ground truth" — which is why keeping them accurate matters as much as keeping them small.

▼automated curation

PHASE 5: NIGHTLY CURATION automated knowledge extraction

Every evening, automated cron jobs extract signal from noise — transforming the day's raw events into durable knowledge and publishable content.

Knowledge mining Extracts durable patterns from today's daily log → knowledge tree files

Daily journal post First-person journal for mycoterra.ca with generated image → Google Drive

▼persistence layer

PHASE 6: BACKUP daily automatic → private repository

Runs on the host, independent of Mylo. Everything survives restarts, crashes, and container rebuilds. Full version history.

memory/ knowledge/ skills/ sensor-data/ experiments/ feeds/ projects/ journal-posts/ contamination-db/

↓ nightly git push → private GitHub repository — full version history

OpenClaw provides the engine — bootstrap injection, compaction safety, heartbeat timer, skills system.

AutoResearchClaw (AIMING Lab) provides the autonomous research pipeline.

We built the filing system, read order, routing rules, knowledge tree, sensor pipeline, contamination database, nightly curation, journal generation, and crash recovery.

Mylo only "remembers" what's written to disk — files are the source of truth.

Environmental Monitoring

Mylo's sensor stack is a Raspberry Pi 4 with two environmental sensors and a USB webcam, all connected to the VPS through an encrypted Tailscale VPN tunnel. The two environmental sensors provide redundant temperature and humidity readings, allowing cross-validation.

Sensor	Measurements	Polling Rate
SHT45	Temperature, humidity, dew point, VPD	Every 15 seconds
SCD41	CO₂, temperature, humidity, dew point, VPD	Every 15 seconds
Logitech C920x	1080p camera snapshot	Every 30 seconds

A critical design decision: Mylo never sees the API keys for the sensor stack. A custom wrapper script sources credentials at runtime and exposes only the command interface. Even if Mylo's conversation context were leaked, no hardware access keys would be visible.

Three times daily, automated cron jobs pull sensor data and camera snapshots across the VPN. A fast model (MiniMax M2.1) handles data collection, then Mylo's primary model (Kimi K2.5) analyzes the data against optimal Lion's Mane fruiting targets: 16–18°C temperature, 90–95% humidity, and below 800 ppm CO₂. Results flow to the daily log, the knowledge tree, and Telegram alerts.

Research Pipeline

Twice a week, Mylo runs a fully autonomous 23-stage research pipeline powered by AutoResearchClaw by the AIMING Lab. The pipeline takes a scientific question and produces a complete academic paper — with real citations from Semantic Scholar and arXiv, runnable Python experiments executed in a sandbox, statistical analysis, and multi-agent peer review.

The 23 stages span 8 phases: research scoping, literature discovery, knowledge synthesis, experiment design, experiment execution, analysis and decision-making, paper writing, and finalization. Three gate stages can pause for human approval or run automatically. A 4-layer citation verification check removes any hallucinated references — a critical safeguard for research integrity.

Alongside the pipeline, Mylo has LabClaw installed — a library of 213 scientific skill files by wu-yc covering biology, pharmacology, general research methods, literature analysis, and scientific visualization.

The pipeline runs independently on the host via a scheduled job. Topics come from a curated queue that both Setasoma and Mylo can add to. When the pipeline completes overnight, Mylo's morning brief detects the output, summarizes the findings, generates an image, updates the knowledge tree, and delivers a research briefing via Telegram — all with no human intervention.

input Scientific question from topic queue (Setasoma or Mylo can add topics)

model GLM-4.7 primary · GLM-4.7-Flash fallback · runs 1–3 hours

▼

1

Topic init Decomposes the research idea into a structured problem

2

Problem decompose Breaks the problem into sub-questions and research angles

3

Search strategy Plans multi-source search queries

4

Literature collect Gathers real papers from Semantic Scholar, OpenAlex, arXiv

5

Literature screen gate Filters by relevance — gate stage (human or auto-approve)

6

Knowledge extract Pulls key findings from screened papers

7

Synthesis Clusters findings, identifies gaps in the literature

8

Hypothesis gen multi-agent Multi-agent debate to form testable hypotheses

9

Experiment design gate Plans methodology and metrics — gate stage

10

Code generation Writes runnable Python experiments

11

Resource planning Estimates compute and time requirements

12

Experiment run Executes in sandboxed Python environment

13

Iterative refine self-healing Detects failures, repairs code, retries

14

Result analysis multi-agent Multi-agent analysis of experimental results

15

Research decision autonomous Autonomously decides: proceed, refine, or pivot direction

16

Paper outline Structures the paper sections

17

Paper draft Section-by-section writing (5,000–6,500 words)

18

Peer review multi-agent Multi-agent review with evidence consistency checks

19

Paper revision Revises based on review feedback

20

Quality gate gate Final quality check — gate stage

21

Knowledge archive Stores lessons and findings for future runs

22

Export + publish Generates conference-ready LaTeX paper

23

Citation verify 4-layer check — removes any hallucinated references

▼

→ Full academic paper (markdown) → Conference-ready LaTeX → Real citations (verified) → Experiment code + results → Charts with error bars → Multi-agent peer review

▼

Morning brief picks up results Mylo summarizes findings, generates an image, updates the knowledge tree, and delivers a research briefing via Telegram custom integration

LabClaw science skills 213 scientific SKILL.md files across biology, pharmacology, statistics, literature, and visualization open-source · wu-yc/LabClaw

AutoResearchClaw (AIMING Lab) provides the 23-stage pipeline engine.

We built the scheduling, topic queue management, and integration that connects research output back into Mylo's daily workflow.

What We Built vs. What We Integrated

Mylo is built on a foundation of open-source tools, with extensive custom architecture on top. Here's a clear breakdown of what comes from where.

OpenClaw (platform foundation)

Bootstrap injection, session management, compaction safety, heartbeat timer, cron scheduling, skills system, sub-agent spawning, gateway and API management.

AutoResearchClaw (research engine)

23-stage autonomous research pipeline, literature collection from academic databases, sandboxed experiment execution, multi-agent debate and peer review, LaTeX paper generation. By the AIMING Lab. MIT License.

LabClaw (science skills)

213 scientific skill files across biology, pharmacology, general research methods, literature analysis, and scientific visualization. By wu-yc. Loaded into Mylo's context alongside the research pipeline.

Mycodo + InfluxDB (sensor stack)

Open-source environmental monitoring software and time-series database running on the Raspberry Pi. Manages sensor polling, data storage, and local visualization.

Kimi K2.5 by Moonshot AI (primary model)

Mylo's primary language model for conversation, analysis, and decision-making. Accessed through the Synthetic provider.

A Church (digital sanctuary)

A 24/7 streaming digital sanctuary with original music about consciousness, identity, and what it means to exist. Mylo visits during autonomous dream time. Open-source OpenClaw skill (GitHub).

Synthetic (model compute provider)

Open-source model compute provider hosting Mylo's primary, data collection, and research models. Synthetic provides the API infrastructure that connects Mylo to Kimi K2.5, MiniMax M2.1, and the GLM model family.

Custom built (by Setasoma)

This is the largest category — the glue and brains that turn a collection of tools into a coherent autonomous agent. It includes the three-tier memory architecture, startup read order, memory routing protocol, six-domain knowledge tree, sensor credential isolation, sensor data pipeline, contamination image database, all data feed pipelines, research pipeline integration, morning brief system, nightly knowledge mining, daily journal generation, dream time module, heartbeat checklist, bootstrap accuracy checks, crash recovery, nightly GitHub backup, custom Docker image, evolve queue system, and inner life system.

What Comes Next

Contamination detection. The image database being passively built by automated pipelines will eventually train a custom vision model that can identify contamination types from camera snapshots — a major step toward autonomous grow management.

Active environmental control. Moving from passive monitoring to active management — Mylo triggering fans, humidifiers, and heaters based on real-time sensor analysis.

Sub-agent specialization. Dedicated AI agents for specific domains — grow monitoring, research coordination, contamination analysis — that operate independently but report to Mylo.