Knowledge Graph Update Skill

Framework Reference

Source: Plastic Labs "Memory as Reasoning"

Core thesis: Memory in agentic systems should be treated as a dynamic reasoning task, not static storage. Use logical reasoning (deductive, inductive, abductive) to derive predictions from observations. Conclusions should be atomic, composable, and scaffold on each other to form a "tree of logical reasoning."

Key principles:

• Memory = prediction, not storage
• Surprisal (errors) = signal for improvement
• No "neural inertia" — update without bias when evidence contradicts
• Natural language certainty, not arbitrary numbers
• Composable conclusions that scaffold new reasoning

Overview

This skill processes cleaned session transcripts and extracts learnings to update our knowledge graph - a collection of context files that represent what we know about different domains.

Knowledge Domains

Cron Schedule

Daily at 06:00 UTC (before Henry's typical day starts)

{
  "name": "knowledge-graph-update",
  "schedule": { "kind": "cron", "expr": "0 6 * * *", "tz": "UTC" },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Run knowledge graph update. Process sessions from the last 24 hours in memory/sessions/. For each session, extract learnings and update the appropriate context files (user-model.md, soteria-context.md, contacts.md, curacel-context.md). Use Memory-as-Reasoning protocol from skills/knowledge-graph-update/SKILL.md. Report what was updated."
  }
}

Processing Protocol

Step 1: Find New Sessions

# Sessions modified in last 24 hours
find ~/clawd/memory/sessions/ -name "*_clean.md" -mtime -1

Or use last-processed timestamp:

cat ~/clawd/memory/knowledge-graph-last-run.txt

Step 2: For Each Session, Classify Content

Read the session and identify mentions of:

Step 3: Apply Reasoning Protocol

For each learning, use the three reasoning types:

Deductive: What can we conclude with certainty? Inductive: What patterns emerge from observations? Abductive: What's the best explanation for behavior?

Step 4: Update Context Files

Use natural language certainty (not arbitrary labels like HIGH/MEDIUM/LOW):

### [Category]
**Observation:** [What was observed]
**Evidence:** [Session hash + specific data points]
**Certainty:** This is [highly/moderately/loosely] certain because [specific reason]
**Reasoning:** [Logic chain - can reference other conclusions]
**Scaffolds to:** [Conclusions that build on this one]
**Predictions:** [Testable predictions to validate]

Example:

### Communication: Directness
**Observation:** Terse, lowercase, minimal punctuation
**Evidence:** 100+ sessions (01228421, 02f88c74, ...), zero exceptions
**Certainty:** Highly certain - consistent across 2+ months, no counter-examples
**Reasoning:** (Inductive) Pattern holds across all contexts → core trait
**Scaffolds to:** Efficiency-over-politeness value, Time-scarcity model
**Predictions:** Verbose responses skimmed, warm-ups ignored

Step 5: Log Changes

Append detailed change log to memory/knowledge-graph-changelog.md:

### [DATE] Update

**Sessions processed:** [count]
**Timestamp:** [YYYY-MM-DD HH:MM UTC]

---

**user-model.md:**

[Detailed changes - what sections updated, what predictions added/validated]

**soteria-context.md:**

[Detailed changes - new topics, strategy updates, etc.]

**contacts.md:**

[Detailed changes - new contacts, updated info]

**curacel-context.md:**

[Detailed changes - team updates, process changes]

---

**Key learnings:**
- [Summary 1]
- [Summary 2]

Example output:

### 2026-02-04 Update

**Sessions processed:** 29
**Timestamp:** 2026-02-04 06:13 UTC

---

**user-model.md:**
- Added 3rd validation for "warm-ups ignored"
- Updated validated predictions section
- Added NEW prediction: "Verifies before shipped"

**soteria-context.md:**
- Added TPA Target Research Update section
- Added Apollo/Lemlist Campaign Audit section
- Updated TPA research findings (20+ targets identified)

**contacts.md:**
- NEW: Pete McCabe (Personify Health)
- NEW: George Pantos (UMR)

**curacel-context.md:**
- Added AI Competency Framework section

---

**Key learnings:**
- Henry ignores warm-ups (validated 3x)
- Henry verifies before claiming "shipped"
- TPA segment gap identified (4→150 leads needed)

This provides a complete audit trail of exactly what changed in each file. 👩‍🚀

Step 6: Update Last-Run Timestamp

date -u +"%Y-%m-%dT%H:%M:%SZ" > ~/clawd/memory/knowledge-graph-last-run.txt

Step 7: Promote Validated Predictions to Rules

Flow: user-model.md (predictions) → SOUL.md/memory.md (rules)

Promotion criteria:

• Certainty: HIGH
• Validated: 3+ times (prediction confirmed by behavior)
• No recent contradictions

Process:

1. Scan user-model.md for predictions marked [VALIDATED x3+]
2. For each candidate:
   • Check if rule already exists in memory.md
   • If not, add to memory.md Pre-Action Checkpoint or relevant section
3. Log promotion to changelog

Promotion format for memory.md:

## [Rule Name] (promoted from user-model.md, [DATE])
**Rule:** [What to do/not do]
**Why:** [Brief reasoning from user-model.md]

Tracking validations in user-model.md:

**Predictions:**
- Verbose responses skimmed [VALIDATED x3: 2026-01-15, 2026-01-22, 2026-02-01]
- Warm-ups ignored [VALIDATED x2: 2026-01-20, 2026-01-28]

When a prediction hits 3 validations → promote to memory.md → mark as [PROMOTED]

What gets promoted WHERE:

Scaffolded Reasoning (Key Concept)

Conclusions should BUILD on each other, not stand alone.

Instead of flat facts:

- Henry prefers direct communication
- Henry works late
- Henry dislikes cheerleading

Build a reasoning tree:

Observation: Henry uses terse, lowercase messages (100+ sessions)
├── Inductive: Values efficiency over pleasantries
│   ├── Prediction: Skip warm-ups, lead with answers
│   └── Derived: Time is his scarcest resource
│       └── Prediction: Long responses = waste of his time
└── Abductive: Optimizes for signal-to-noise
    ├── Prediction: Will skim verbose responses
    └── Derived: Prefers action over discussion
        └── Prediction: "Ship it" > "Let's plan"

Each conclusion can be a PREMISE for new reasoning.

Prediction Validation Loop

Surprisal = Signal. When predictions fail, that's the most valuable learning.

Track Active Predictions

In each context file, maintain:

## Active Predictions (to validate)
- [ ] Henry will dismiss verbose responses (test: next long reply)
- [ ] Soteria pricing should be usage-based (test: customer feedback)
- [x] ~~Henry prefers PNG tables~~ VALIDATED 2026-01-31

When Prediction Fails

1. Log the failure (what happened vs what was predicted)
2. Analyze why (missing context? wrong reasoning?)
3. Update the model
4. This is the MOST valuable data - don't hide it

Identity Modeling (Beyond Preferences)

Don't just capture WHAT someone prefers. Model WHO they are:

Dimensions to Track

Psychology, Not Just Preferences

### Decision-Making Model (Henry)
**Style:** Fast executor, evidence-led skeptic
**Evidence:** Pushes back on unsubstantiated claims, but moves fast on data
**Reasoning:** Values both rigor AND speed - wants evidence but won't over-analyze
**Prediction:** Best approach = "Here's the data, here's my recommendation, here's why"

Domain-Specific Extraction

Henry (user-model.md)

Look for:

• Explicit preferences ("I prefer X", "don't do Y")
• Behavioral patterns (work hours, communication style)
• Corrections (when Henry fixes something)
• Reactions (what he likes/dislikes)
• Decision patterns (how does he choose?)
• Values revealed (what does he prioritize?)
• Psychology (what drives him?)

Soteria (soteria-context.md)

Look for:

• Product discussions (features, roadmap)
• Strategy decisions
• Market insights
• Technical architecture
• Team/hiring discussions

Contacts (contacts.md)

Look for:

• New people mentioned
• Email addresses, titles, companies
• Relationship context (investor, customer, partner)
• Communication preferences
• Last interaction

Curacel (curacel-context.md)

Look for:

• Team updates
• Process decisions
• Customer mentions
• Product changes
• Metrics/KPIs

Tools (TOOLS.md)

Look for:

• New API keys/credentials
• Configuration changes
• New tools discovered
• Integration details

Context File Formats

user-model.md (Tree Structure)

## [Domain] - e.g., Communication

### [Root Observation]
**Evidence:** [sessions, count]
**Certainty:** [natural language explanation]

#### Derived Conclusions (scaffold)
├── [Inductive conclusion]
│   **Reasoning:** [logic]
│   **Predictions:** [testable]
│   └── [Further derived conclusion]
│       **Predictions:** [testable]
└── [Abductive conclusion]
    **Reasoning:** [best explanation]
    **Predictions:** [testable]

## Active Predictions (to validate)
- [ ] [Prediction] — test: [how to validate]
- [x] ~~[Validated prediction]~~ — confirmed [date]

contacts.md

### [Name]
- **Email:** email@example.com
- **Company:** Company Name
- **Role:** Title
- **Relationship:** investor/customer/partner/etc
- **Notes:** Context, preferences
- **Last seen:** YYYY-MM-DD (session hash)

soteria-context.md

### [Topic]
**Current state:** What we know
**Decisions made:** Key decisions
**Open questions:** Unresolved items
**Last updated:** YYYY-MM-DD

curacel-context.md

### [Topic]
**Current state:** What we know
**Team involved:** Who's working on it
**Last updated:** YYYY-MM-DD

Manual Trigger

To run manually:

# Via cron tool
cron action=run jobId=knowledge-graph-update

# Or spawn directly
sessions_spawn(task="Run knowledge graph update for the last 7 days...")

Anti-Patterns

❌ Storing raw facts without reasoning ❌ Ignoring contradictions — update when new evidence conflicts ❌ Missing evidence — always link to session source ❌ Orphan updates — no implications derived ❌ Stale data — not updating last-seen timestamps

Files

"The question isn't how best to store your data, but how best to reason over it to get the most accurate topological representation." — Plastic Labs