# Personal Knowledge Companion## ConceptA persistent knowledge layer across all agents that maps what you know, what you're learning, and where your gaps are — then tells you about those gaps proactively.The visual/exploration layer is a **knowledge graph** — relationships and connections, not just stored documents. This makes gap detection natural: missing edges in the graph *are* the gaps.Domain separation:- 🏫 **School** — 2890 bot (robotics team)- 🏭 **Work** — PSB bots (brewery ops)- 🎮 **Play** — crash-bot / HHS-Hackers crew## Core InsightCurrent agents are mostly command-and-reply. They don't do research, self-direct, or maintain persistent context between sessions. They feel like fancy autocomplete, not assistants.The question: what would make them feel less like tools and more like... something with continuity?## Architecture### The Silo ProblemThree agent domains are intentionally isolated:- 🏫 School (2890) — student data, competition strategy- 🏭 Work (PSB) — brewery ops, business data- 🎮 Play (HHS-Hackers) — crew knowledge, projectsThe wiki needs to cross domains without breaking the walls. Solution: a librarian agent.### Multi-Person ScalingSame shared knowledge graph, different personal layers per person:- Each person gets an entity page with confidence scores, scale preferences, learning style- Shared facts (ESP32 specs, VLAN concepts) compound as every person's learning adds resources- Access control maps to existing Discord server structure- Kyle's CEH training → HHS-Hackers server- Bruno's PathPlanner → 2890 server- Your VLAN training → DMs or HHS-Hackers- The compounding effect: every person's learning path makes the knowledge graph richer for the next person### Professor Agent (The Teacher)Not just a librarian — a **teacher** that:1. Sees what you're working on (Pi-hole setup → VLANs coming next)2. Assesses where you are ("almost none" on VLANs)3. Finds resources matched to your scale (home lab, not enterprise)4. Organizes into a learning path (training thread under VLAN)5. Delivers proactively without being askedThe difference: A librarian says "here's everything filed under VLAN." A professor says "here's what you need right now, at your level, for your situation."**How it works with the wiki stack:**- Project entities have required skill tags (Pi-hole → DNS, networking, VLANs)- Person entities have confidence scores per skill (VLANs: 0.1, never touched)- Professor maps required skills against known skills → finds gaps- Searches curated sources scoped to your context (home-lab scale, not enterprise)- Delivers as a Discord training thread, which itself becomes wiki content- When you come back and say "I get VLANs now", professor updates the claim- Every person's learning path compounds into the shared knowledge graph**Context-aware filtering** — the professor knows your scale:- `bestUsedFor: home-lab, small-scale, hands-on learning`- `notEnoughFor: enterprise, production, large-scale`- This turns "47 VLAN resources" into "3 that match your setup"### Vault StructureEach agent owns its own vault partition. Professor reads metadata across partitions and writes to a shared cross-reference layer. Walls stay up.### Layers1. **memory-wiki** — Storage layer. Bridge mode. Obsidian-compatible. Structured claims with provenance and confidence.2. **Knowledge graph** — The exploration/visual layer. Maps relationships and connections. Missing edges = gaps.3. **Gap detection** — Runs on the professor. Proactive intelligence. Surfaces what's stale, what's missing, what's the natural next step.4. **Professor agent** — The teacher. Cross-domain reader, context-aware resource finder, learning path organizer. Bridges the silos without breaking them.Build order: wiki config → professor agent → gap detection → knowledge graph visualization### Exec Power Constraint- **crash-bot-DM (this agent, -topher's DMs):** HAS exec. Can build, wire, and touch the system.- **crash-bot-public (server channels):** No exec. Config changes must be drafted here and applied by -topher or crash-bot-DM.- **crash-bot-public exec** will be gated via `exec-approvals.json` allowlist + ask mode — commands require -topher's approval via `/approve`.## Exec Approval SetupSee: `projects/exec-approval-setup.md`**Status:** Partially implemented. Waiting for gateway reboot to complete configuration.### What's done:- [x] Backed up `openclaw.json` → `openclaw.json.pre-exec-approval-20260502-120941`- [x] Backed up `exec-approvals.json` → `exec-approvals.json.pre-setup-20260502-120941`- [x] Updated `exec-approvals.json` with full policy (crash-bot-public: allowlist+ask, crash-bot: pre-approved read commands)### What's blocked:- [ ] Update `openclaw.json` agent tool configs for exec settings- [ ] Restart gateway and verify exec works- [ ] Test approval flow**Blocker:** Gateway rejecting all exec with "pairing required" — full system reboot needed## What's MissingThe "no research" problem. Agents respond when called but don't:- Proactively gather information- Maintain long-term context outside of session- Work on background tasks between interactions## Memory FileSee: `memory/personal-knowledge-companion.md` — full concept, references, and status## Status- [x] Concept documented- [x] Discord channel (#personal-knowledge-companion)- [x] Knowledge graph (not Empire visual) as exploration layer- [x] Obsidian as client for vault browsing/graph- [x] memory-wiki plugin research complete (bridge mode, Obsidian render)- [x] Professor agent vision documented (teacher, not librarian; multi-person scaling)- [x] Multi-person scaling architecture documented- [x] Exec approval setup documented (projects/exec-approval-setup.md)- [ ] exec enabled for crash-bot-public (awaiting crash-bot-DM implementation)- [ ] memory-wiki plugin enabled and configured- [ ] Vault structure created (entities, concepts, syntheses, sources, reports)- [ ] Bridge mode connected to active memory- [ ] Professor agent- [ ] Gap detection logic- [ ] Knowledge graph visualization (Obsidian graph view)## memory-wiki Plugin Research- Built into OpenClaw, no extra install needed- Vault modes: isolated (own vault), bridge (reads active memory artifacts), unsafe-local (escape hatch)- Recommended: **bridge mode** — QMD for recall, wiki for synthesized knowledge- Vault layout: entities/, concepts/, syntheses/, sources/, reports/, _views/- Obsidian-compatible render mode built in- Structured claims with confidence, provenance, evidence- Auto-generates dashboards: open-questions, contradictions, low-confidence, stale-pages, relationship-graph- Tools: wiki_search, wiki_get, wiki_apply, wiki_lint, wiki_status- Config path: plugins.entries.memory-wiki.config- Vault default path: ~/.openclaw/wiki/main- Bridge mode indexes: memory artifacts, dream reports, daily notes, memory root files, memory events### Hardware ConstraintsRunning on Dell Aurora-R3 (i7-2600K, 32GB DDR3). File-based wiki = zero compute cost. Obsidian = client-side only, no server. Qdrant already running. Gap detection = lightweight heartbeat task. No roadblocks on the core build — only a real-time interactive graph server would strain this box, and Obsidian's local graph view sidesteps that entirely.## Related- Claw Empire (GitHub: GreenSheep01201/claw-empire) — separate project for later. Different problem space (orchestration/workflow), not the visual layer for this.- Mission Control — already has agent topology, could be foundation