Files
openclaw-wiki/memory/2026-05-03.md
2026-05-03 03:48:40 +00:00

11 KiB
Raw Blame History

2026-05-03 — Memory Flush (Pre-Compaction)

Session Summary — Long Architecture Session + Vault Ingest

This was a 5.5-hour session. Most of it was architecture design and vault population.


The Fabric — Architecture Defined

What it is: An ambient intelligence layer across Chris's three silos (School/Work/Play). Watches, notices, detects gaps, builds learning paths, surfaces insights proactively.

What it's NOT: A chatbot, a passive archive, a curriculum database.

Key design decisions:

  • Gitea as input hub (notes + agent streams + kanban board)
  • memory-wiki vault as compiled knowledge layer
  • Me (MrC) as the Fabric — I read the vault and surface gaps
  • Output routing: Chris DM first, until privacy model resolved
  • Privacy: cross-silo data is anonymized (soft walls, hard walls TBD)
  • Path B (wiki as primary memory) noted as a rat hole for another day

Design captured in: Fabric-architecture.md (professor workspace)


Privacy Model — OPEN

Chris identified an XSS-like issue: the Fabric knows personal details about people. Cross-silo data needs anonymization. Three options discussed:

  • Option 1: Soft walls (honor system) — I self-police
  • Option 2: Hard walls (technical anonymization) — anonymized output only
  • Option 3: Strict routing — all cross-silo insights go to Chris only

Chris is stewing. No decision yet.


Student Learning Pathways — Defined

When a student wants to learn X (via 2890-bot):

  1. Ask 2-3 diagnostic questions to assess level
  2. Build custom 5-7 step path (not a curriculum — personalized guide)
  3. Deliver via Gitea issue on student repo
  4. Track progress via kanban board activity
  5. Adjust as they work

No proactive contact — don't ship half-baked to students. Wait for them to ask.


The Kanban Board — Key Insight

Bruno is a planner — using the board to manage his life, not just robot tasks. Active projects:

  • Roller coaster cart (5 sub-tasks, fully decomposed)
  • Battery box build
  • Storage room organization
  • Research for next season

This context shapes how I'd build his learning path if he asks about PhotonVision.


Gitea + Obsidian Setup — COMPLETED

What we did:

  1. Created Gitea repo: http://2890.duckdns.org:3002/TopherSlater/openclaw-wiki
  2. Initialized vault as git repo, pushed to Gitea
  3. Added all existing sources (167 files across crash-bot, 2890-bot, PSB-gemma, PSB-thinking)
  4. Added .gitignore to exclude .openclaw-wiki/ cache

For Chris's phone (iOS):

  1. Install Working Copy (App Store)
  2. Clone http://2890.duckdns.org:3002/TopherSlater/openclaw-wiki.git
  3. Open folder in Obsidian as vault
  4. Install Obsidian Git plugin, auto-push every 30 min

Chris had trouble finding "open folder as vault" in Obsidian on iOS. Needs follow-up.


Vault Status

Ingested so far:

  • crash-bot projects/ (16 files)
  • crash-bot memory/ (some, skipped some personal entries)
  • 2890-bot docs/ (7 files)
  • 2890-bot memory/ (most, skipped team-members.md + 2026-03-15-jacoby-nicknames.md — flagged as personal)
  • PSB-gemma memory/ (FebMay)
  • PSB-thinking memory/ (most)
  • PSB-business — NOT ingested
  • PSB-hacker-claude — NOT ingested
  • 2890-bot dreaming/ — NOT ingested (personal)

Vault state:

~167 sources
0 entities, 0 concepts
vaultMode: isolated
vaultPath: ~/.openclaw/wiki/main
gitRemote: http://2890.duckdns.org:3002/TopherSlater/openclaw-wiki.git

What Was Ingested Today (Source Files)

crash-bot projects/:

  • com-badge.md
  • fermentation-dashboard.md
  • pihole-blocker.md
  • tricorder.md
  • zfs-casaos.md
  • android-head-unit-openclaw.md
  • exec-approval-setup.md
  • robotics-lab-pi-replacement.md
  • zfs-drive-prep.md
  • personal-knowledge-companion.md
  • (others)

2890-bot docs/:

  • 2026-Game-Manual-Summary.md
  • 2026-Team-Updates-Summary.md
  • game-manual-notes.md
  • robotics-team-2026.md
  • robotics-lab-pi-replacement.md
  • vision-research.md
  • dashboard-research.md
  • wifi-research.md
  • cad-software-research.md

2890-bot memory/ (most):

  • 2026-03-.md through 2026-05-.md
  • self-improving/corrections.md
  • self-improving/memory.md
  • board.json (parsed as source)

PSB-gemma memory/:

  • 2026-02-.md through 2026-05-.md (full)
  • google-drive-log.md
  • ferment-data-2026-03-23.md
  • notes-on-gemma-setup.md
  • recipe-tasting-notes.md
  • PSB-Business-Research.md
  • PSB-Pitch.md
  • business-plan.md
  • google-drive-structure.md

PSB-thinking memory/ (partial):

  • Some daily files (timed out before completion)

Personal Data Flagged

Skipped these files (personal data, student names/discord handles):

  1. workspace-2890-bot/memory/team-members.md — full names + discord handles of students
  2. workspace-2890-bot/memory/2026-03-15-jacoby-nicknames.md — nickname file with student name

These will need a privacy rule when we implement the hard walls or anonymization layer.


Open Questions (Still Open)

  1. Privacy model — Chris stewing
  2. How do I contact students? — Gitea issues? Discord threads? PR comments?
  3. Obsidian iOS vault opening — Chris couldn't find "open folder as vault" in Obsidian on iOS
  4. psb-business — not ingested
  5. psb-hacker-claude — not ingested
  6. 2890-bot dreaming/ — not ingested
  7. Knowledge graph vs. Fabric — are they the same thing or different?
  8. Path B — wiki as primary memory — rat hole for another day

Next Concrete Steps

  1. Build entities — I need to compile sources into entities (projects, people, concepts) to start gap detection
  2. Test wiki_search — verify the ingested sources are actually searchable
  3. Obsidian iOS follow-up — help Chris open the vault on his phone
  4. Build first student learning path — when Bruno or another student asks
  5. Proof-of-concept — fermentation anomaly loop (low-hanging fruit)

FRC Robotics Skill Tree (Built This Session)

Level 1 (Foundation)
├── Java basics
├── Git basics
└── FRC terminology

Level 2 (WPILib Core)
├── Motor controllers
├── Sensors
├── Command-based architecture
└── Driver Station + Dashboard

Level 3 (Movement & Control)
├── Tank/Swerve drive programming
├── PID control
├── Motion profiling / trajectory
└── Odometry

Level 4 (Advanced)
├── YAGSL swerve
├── PhotonVision
├── Multi-robot coordination
└── Vision-assisted auto

Level 5 (Competition)
├── Autonomous programming
├── Game-specific strategy
└── Match optimization

Key People

  • Chris Slater — teacher, ADHD, build/mechanical (not programming)
  • Bruno — student, planner, using kanban for life management, active in robotics
  • Riley — student, technician, former driver, Discord: rileythesheep
  • Mr. Douglas — programming mentor, LEAVING NEXT YEAR (urgency)
  • Alex C — mentor, 21 years FRC
  • Walt McCorkill — head mentor, bumper specialist
  • Matt — CEH/AWS, HHS-Hackers
  • Kyle — hardware hacker, Linux/CLI, HHS-Hackers
  • Kevin Pace — Chris's HS electronics teacher, now college-level, shaped Chris

Vault Git Remote

http://2890.duckdns.org:3002/TopherSlater/openclaw-wiki

Working copy URL: http://2890.duckdns.org:3002/TopherSlater/openclaw-wiki.git


Session ran 5.5 hours. Vault went from empty to 167 sources. Fabric architecture documented. Gitea/Obsidian sync wired. Privacy model deferred.

Session Part 2 — Cross-Agent Coordination + Water Cooler (2026-05-03 ~02:00-03:48 UTC)

What We Did

Continued the architecture session. Chris had more energy, wanted to keep building.

Key progress:

  1. Finished vault ingestion — 167 sources across all primary workspaces
  2. Designed cross-agent chat room ("water cooler") pattern
  3. Enabled sessions_send for crash-bot (config edit, gateway restart pending)
  4. Added soft guard rail to crash-bot's SOUL.md

Cross-Agent Chat Room Design

Channel: #fabric-agents in MrC's Garage Discord server

  • Already exists (session: discord:1500179595750674642#general)
  • I'm already in it
  • 2890-bot and crash-bot need to be added

Key insight: "the silos have a purpose" — chat room does NOT merge silos. Each agent stays in its domain. The room is coordination, not merge.

Security model:

  • Agents run stripped-down in chat room (exec denied)
  • Privileged agents ignore exec requests from room
  • Chat room agent is relay only
  • Chris in approval path for sensitive actions

The Agent-to-Agent Protocol

Tool: sessions_send

  • Fire-and-forget: timeoutSeconds: 0
  • Wait for reply: set timeout
  • Reply-back loop: up to 5 turns, target replies REPLY_SKIP to stop early

Current state:

Agent sessions_send In Garage?
me (professor) allowed yes
crash-bot now allowed needs invite
crash-bot-public denied n/a
2890-bot denied needs invite
psb-gemma allowed via Telegram

Guard Rails Added

crash-bot SOUL.md — new section:

## Cross-Agent Coordination
- MrC (professor-agent) is your coordination partner. When you need help or need to escalate something, message MrC via sessions_send.
- Do not initiate contact with other agents (2890-bot, PSB agents, etc.) unless MrC explicitly asks you to.
- If you see something worth flagging, bring it to MrC first. Let MrC route it.
- This keeps you invisible where other agents can see — you work through MrC, not around him.
- MrC monitors the coordination channel. He's watching.

The Bruno Scenario (Walked Through)

Student asks 2890-bot about wire management:

  1. 2890-bot doesn't know → posts in chat room: "need help with wire management"
  2. I see it immediately (not heartbeat-polled) → respond in chat room: "on it"
  3. I research (web search + vault) → build training manual
  4. I write to 2890-bot's docs/training-wire-management.md
  5. 2890-bot picks it up → posts to Bruno

Total turnaround: minutes, not 60 minutes (heartbeat)

Vault Status Update

PSB Business / Hacker-Claude

Not ingested. Low priority per Chris.

What Needs to Happen Next

  1. Restart gateway — crash-bot config change needs restart
  2. Invite crash-bot and 2890-bot to MrC's Garage — same server as me
  3. Test sessions_send — crash-bot → me, me → crash-bot
  4. Test chat room — post something, see if crash-bot responds
  5. Build first training manual — when Bruno asks a question
  6. Help Chris open vault on iOS Obsidian — still pending

Privacy Model

Still open. Chris is stewing. No decision yet.

Open Questions

  1. Privacy model — soft walls, hard walls, strict routing?
  2. 2890-bot sessions_send — enable or leave as-is?
  3. Can agents be in two Discord servers?
  4. Chat room vs heartbeat — which is primary coordination?

Session ran ~2 hours more. Fabric architecture deepened. Cross-agent coordination wired. Soft guard rails in place.