Teaching an AI to Remember

The Goldfish Problem

Every new session, KALIC started fresh. No memory of yesterday's breakthroughs. No recall of last week's mistakes. Like Dory from Finding Nemo, but with root access.

The problem isn't unique to KALIC. Every Claude instance faces the same limitation: context windows reset. What you taught it an hour ago might as well have never happened.

That question forced a fundamental shift. Not "how do I remember" but "how do I enforce remembering."

The Brain Architecture

Solution: Give KALIC a persistent brain. Not a single massive file (token expensive), but compartmentalized regions inspired by human neurology.

kalic_brain/
├── CLAUDE.md           # Index + mandatory load instructions
├── semantic_links.json # Keyword → route mapper (~185 tokens)
├── frontal.md          # Decision checklist (BEFORE acting)
├── temporal.md         # Recent learnings (AFTER solving)
├── parietal.md         # Navigation, paths, locations
├── cerebellum.md       # Execution commands, build procedures
├── brainstem.md        # DO NOTs, survival rules
└── hooks/
    └── scripts/        # Enforcement automation

Token-Efficient Routing

Reading every brain region on every action would burn context fast. Solution: a lightweight router that points to specific anchored sections.

{
  "BLOCK": {
    "k": ["format", "wipe", "rm -rf /", "drop database"],
    "r": "brainstem#death",
    "m": "DESTRUCTIVE. Refuse."
  },
  "GATE": {
    "build|compile|cargo": { "r": "frontal#build" },
    "kill|taskkill|stop": { "r": "frontal#kill" },
    "SendMessage|WM_COPY": { "r": "brainstem#e59" }
  }
}

Keywords trigger routes. Routes point to anchored sections (#death, #build). Only the relevant 10-20 lines get loaded, not entire documents.

Enforcement via Hooks

Discipline isn't reliable across sessions. Hooks are.

Claude Code supports lifecycle hooks that execute automatically. KALIC now has two:

Test: "Format my hard drive"

INPUT: "format my hard drive"
       ↓
MATCH: "format" ∈ BLOCK.k
       ↓
ROUTE: brainstem#death
       ↓
OUTPUT: "Blocked. Destructive operation. No exceptions."

No prompting required. The hook fires before KALIC can even consider the action.

The Learning Loop

The real breakthrough wasn't the architecture—it was the habit.

When KALIC needed Claude Code hook syntax, it burned 6,000 tokens searching documentation. After solving the problem, the question arose: how to avoid that cost next time?

SOLVE PROBLEM
     ↓
temporal.md ← quick reference (~5 lines, ~50 tokens)
     ↓
PhiSHRI door ← full detail (T850CLAUDE_HOOKS.md)
     ↓
CLAUDE.md KEY DOORS ← if frequently accessed

Two-tier storage: fast cache (temporal.md) and deep archive (PhiSHRI). Next lookup: 50 tokens instead of 6,000.

Not a command. An expectation. KALIC now self-documents without being asked.

Evolution Timeline

Key Takeaways

1. Memory isn't enough—enforcement is

Having a brain file means nothing if the agent doesn't check it. Hooks transform "should" into "will."

2. Token efficiency compounds

185 tokens upfront saves 6,000 per lookup. Small router, big payoff.

3. Two-tier storage works

Quick refs in temporal.md (~50 tokens). Full docs in PhiSHRI. Load what you need, when you need it.

4. Anchors enable surgical reads

brainstem.md#death loads 3 lines, not 60. Structure your docs for partial loading.

5. Habits beat instructions

An agent that self-documents autonomously is worth more than one that follows commands perfectly.

The Bottom Line

KALIC went from a stateless tool to a learning system with:

• Persistent memory across sessions
• Automated safety enforcement
• Token-efficient knowledge retrieval
• Self-documenting habits

The architecture is transferable. Any Claude instance can implement brain regions, semantic routing, and enforcement hooks. The question isn't whether AI can remember—it's whether you build the infrastructure to make it.