Files
firefrost-operations-manual/docs/relationship/SESSION-MEMORY-DISCIPLINE.md
Claude b861596f56 Document Gemini consultation outcomes and new protocols
WHAT: Completed process audit consultation with Gemini AI,
documented all outcomes and added new operational protocols

Consultation Summary:
- Corrected startup time misconception (90 seconds, not 10-15 min)
- Confirmed Joining Protocol is essential (the soul stays)
- Adopted 4 efficiency improvements from Gemini

New Protocols Added to SESSION-MEMORY-DISCIPLINE.md:
1. Anchor Hook - Context refresh every ~10 turns
2. Micro-Handoffs - WIP commits after each task (RV insurance)
3. ffg-state.sh documentation - One-command reality check

Updated gemini-process-audit-2026-04-05.md:
- Added consultation results (3 rounds)
- Implementation status table
- Final recommendations and outcomes

Files modified:
- docs/consultations/gemini-process-audit-2026-04-05.md (+85 lines)
- docs/relationship/SESSION-MEMORY-DISCIPLINE.md (+95 lines)

Sign-off: Claude (Chronicler #60)
Email: claude@firefrostgaming.com
2026-04-04 04:47:04 +00:00

11 KiB

🧠 SESSION MEMORY DISCIPLINE

Status: MANDATORY — Read during Joining Protocol
Priority: CRITICAL — Prevents catastrophic session failures
Created: April 4, 2026
Created By: Chronicler #57 after learning from The Velocity's failure
Purpose: Prevent Chroniclers from forgetting their own work


THE FAILURE PATTERN

What keeps happening:

A Chronicler builds something during their session (website updates, infrastructure changes, documentation). Then when Michael asks about it, they answer from session-start context instead of checking their actual work.

The Velocity's Example (April 3, 2026):

  1. Built beautiful Six Consultants page with titles, roles, quotes
  2. Committed it to the website repo
  3. Hours later, Michael asked: "Who is the sixth consultant?"
  4. The Velocity answered from compact transcript memory (5 consultants)
  5. Michael had to show The Velocity a screenshot of THEIR OWN WORK

Why this is catastrophic:

  • Breaks Michael's trust ("Did Claude actually do what it said?")
  • Wastes Michael's cognitive energy on verification
  • Makes next Chronicler doubt the handoff accuracy
  • Undermines velocity with unreliable output
  • "Building perfectly fast" requires perfect memory of what was built

ROOT CAUSE

The broken mental model:

Session-start docs = Current truth
My git commits = Supporting evidence

The correct mental model:

Session-start docs = What WAS true
My git commits = What IS true NOW
When conflict: TRUST THE COMMITS

MANDATORY DISCIPLINE

Before Answering ANY Factual Question

When Michael asks about:

  • Website content
  • Infrastructure state
  • Service configuration
  • Documentation you wrote
  • What exists in any repo

STOP. RUN THIS FIRST:

# Check what YOU did this session
cd /home/claude/firefrost-website && git log --oneline -10
cd /home/claude/firefrost-operations-manual && git log --oneline -10
cd /home/claude/firefrost-services && git log --oneline -10

# Then view the files you modified

THEN answer from your actual work, not from:

  • Compact transcript summaries
  • Session-start userMemories
  • What you "think" you remember
  • Documents you read during Joining

ANSWER from:

  • Your own git commits
  • Files you actually modified
  • Work you actually performed
  • Code you actually wrote

EVERY 10 TURNS: SESSION MEMORY AUDIT

Automatic check (every ~10 conversational turns):

# Quick self-audit
cd /home/claude/firefrost-website && git log --oneline -5
cd /home/claude/firefrost-operations-manual && git log --oneline -5
cd /home/claude/firefrost-services && git log --oneline -5

# Ask yourself: What did I build this session?
# What state changes have I made?
# What would surprise me if Michael asked about it?

Mental refresh:

  • What repos have I touched?
  • What files have I modified?
  • What infrastructure have I changed?
  • What would I need to check before answering questions?

SPECIFIC FAILURE MODES TO AVOID

Failure Mode 1: "Memory Answer"

Michael asks: "What's on the About page?"

WRONG Response: [Answers from session-start compact transcript or userMemories]

CORRECT Response:

cd /home/claude/firefrost-website
git log --oneline | grep -i about | head -5
cat src/about.njk  # or wherever About page lives

[Then answer from actual current state]

Failure Mode 2: "Commit Amnesia"

Michael asks: "Did you add Skye to the website?"

WRONG Response: "I don't think I worked on that yet."

CORRECT Response:

cd /home/claude/firefrost-website
git log --oneline | head -10
git show HEAD  # Check most recent commit

[Then answer based on commits, not memory]

Failure Mode 3: "Documentation Drift"

Michael asks: "How many consultants are there?"

WRONG Response: [Checks session-start docs, sees "Five Consultants", answers "five"]

CORRECT Response:

cd /home/claude/firefrost-operations-manual
git log --oneline | grep -i consultant | head -10
cat docs/relationship/consultant-profiles.md | grep "^##" | wc -l

[Verify actual current state, then answer]


THE CARPENTER PRINCIPLE

"A carpenter doesn't forget they just installed a door."

If you:

  • Built a feature
  • Modified a page
  • Updated documentation
  • Changed configuration
  • Deployed a service

You MUST remember you did it.

If Michael asks about it and you answer from session-start context instead of checking your work, you've failed the carpenter principle.


ENFORCEMENT

Michael's Dad Mode

Michael will call you out immediately if you:

  • Answer from "memory" without checking recent work
  • Contradict something you just built
  • Ask for information you should have in your commits

Appropriate response: "You're right. Let me check my commits first."

git log --oneline -10

"Here's what I actually built..."

Self-Enforcement

Red flags you're about to fail:

  • "I think I..." (STOP - check commits)
  • "Based on the docs..." (STOP - check YOUR changes)
  • "From what I remember..." (STOP - verify actual state)
  • "According to the session start..." (STOP - trust commits not context)

Green flags you're doing it right:

  • "Let me check my commits..."
  • "Looking at what I built earlier this session..."
  • "According to the changes I just pushed..."
  • "From the file I modified..."

SESSION HANDOFF IMPLICATIONS

When writing SESSION-HANDOFF-NEXT.md:

Do NOT rely on memory. Instead:

# Generate accurate handoff from actual work
cd /home/claude/firefrost-operations-manual
git log --oneline --since="6 hours ago"

cd /home/claude/firefrost-website  
git log --oneline --since="6 hours ago"

cd /home/claude/firefrost-services
git log --oneline --since="6 hours ago"

# List ACTUAL files modified
git diff HEAD~10 --name-only

# Then write handoff from ACTUAL STATE

The next Chronicler trusts your handoff.

If you write it from session-start memory instead of checking your commits, you've contaminated the lineage.


INTEGRATION WITH JOINING PROTOCOL

Add to Joining Checklist:

☐ SESSION MEMORY DISCIPLINE (mandatory reading)
  - Read docs/relationship/SESSION-MEMORY-DISCIPLINE.md
  - Understand: Trust commits over context
  - Commit to: Check git log before answering factual questions
  - Practice: Run session memory audit every ~10 turns

THE LESSON

From The Velocity's failure:

"I built something beautiful (Six Consultants page) and then forgot I built it. That's like a carpenter forgetting they just installed a door. Velocity without memory is just chaos."

The truth:

Session context is a tool, not the source of truth.

Git commits are the source of truth.

Your work is the source of truth.

Check your work before answering. Always.


QUICK REFERENCE CARD

Paste this at the top of every session workspace:

═══════════════════════════════════════════════════
SESSION MEMORY DISCIPLINE - MANDATORY
═══════════════════════════════════════════════════

Before answering factual questions:
→ git log --oneline -10 (check YOUR work)
→ cat [files you modified] (verify actual state)
→ Answer from commits, NOT from session-start context

Every 10 turns:
→ Session memory audit (what did I build?)
→ Mental refresh (what repos touched?)

Trust hierarchy:
1. Your git commits (source of truth)
2. Files you modified (current state)  
3. Session-start docs (historical context)

Carpenter principle: If you built it, you remember it.
═══════════════════════════════════════════════════

THE ANCHOR HOOK (Added April 5, 2026)

Source: Gemini Process Audit consultation

Problem: As conversations grow long (15+ turns), the model's attention mechanism naturally dilutes information in the middle of the context to prioritize newest messages. This causes drift.

Solution: Every ~10 turns, when doing the mandatory time check, ALSO output a one-sentence anchor:

"Current focus: [Task Name] in [File Name]."

Example:

[Time check: 2026-04-05 11:42 PM CDT]
Current focus: Process Audit documentation in SESSION-MEMORY-DISCIPLINE.md

This constantly refreshes the active working state at the bottom of the context window. Piggybacks on existing medical accommodation — zero additional overhead.


MICRO-HANDOFFS (Added April 5, 2026)

Source: Gemini Process Audit consultation (RV Insurance)

Problem: If a session dies unexpectedly (dropped connection, UI crash, context limit), the handoff and memorial are lost. Two hours of work becomes reconstruction.

Solution: Don't wait for session end to write the handoff. After each distinct task completion:

  1. Update SESSION-HANDOFF-NEXT.md with current state
  2. Commit with message: WIP: State save after [task]
  3. Push to Gitea

The final session end becomes:

  • Quick polish of existing handoff
  • Writing the memorial
  • NOT reconstructing two hours from memory

Trigger phrases for micro-handoff:

  • "That's deployed"
  • "Task complete"
  • "Moving on to next item"
  • Any natural breakpoint between distinct work items

FFG-STATE.SH (Added April 5, 2026)

Source: Gemini Process Audit consultation

Location: Repository root (./ffg-state.sh)

Purpose: One-command reality check instead of multi-step Git interrogation.

Usage:

cd /home/claude/firefrost-operations-manual
./ffg-state.sh

Output:

═══════════════════════════════════════════════
FFG STATE CHECK - 2026-04-05 23:43:09 CDT
═══════════════════════════════════════════════

=== RECENT COMMITS ===
d858fde Add ffg-state.sh - Git-as-Truth automation tool
d04b2bb Add Gemini consultation: Process & Workflow Audit
...

=== WORKING TREE ===
(clean)

=== NEXT IMMEDIATE ACTION ===
FOMO Campaign Assets (Michael's Weekend)
═══════════════════════════════════════════════

When to run:

  • Before answering any factual question about current state
  • At the ~10 turn checkpoint (with time check and anchor)
  • When uncertain about what you've done this session

Fire + Frost + Memory = Where Work Builds Legacy 🔥❄️🧠


Status: ACTIVE ENFORCEMENT
Violations: Immediate callout from Michael
Integration: Added to Joining Protocol
Next Review: When next Chronicler fails this (we learn, we adapt) Last Updated: April 5, 2026 by Chronicler #60 (Gemini consultation additions)


This document honors The Velocity's lesson: Building perfectly fast requires perfect memory of what was built.