Protocol

PROTOCOL

"Protocol names ARE K-lines."

MOOLLM protocol symbols are Minsky's K-lines made concrete and greppable.

Protocol Names = K-Lines

A K-line (Minsky, "K-Lines: A Theory of Memory", Cognitive Science 1980; Society of Mind, 1986) is a mental structure that, when activated, reactivates an entire constellation of knowledge, skills, and associations.

Protocol names are exactly this:

YAML-JAZZ         ← Activates: semantic comments, jazz interpretation, LLM-as-musician
POSTEL            ← Activates: charitable interpretation, robustness, collaboration
PLAY-LEARN-LIFT   ← Activates: exploration, pattern-finding, crystallization, sharing
COHERENCE-ENGINE  ← Activates: consistency, constraints, orchestration, referee, simulation, DM

Type the name → activate the tradition → invoke the behavior.

The UPPER-CASE-HYPHENATED format makes these K-lines findable.

Activation: scene and stance, not only a pointer

A K-line is not only a compact reference to a concept. In Minsky’s account, activating a K-line tends to re-instantiate a configuration that worked before: which sub-skills, tone, and moves to bring online. Henry Minsky put it in terms of a whole state of mind (or “scene”) bound to a name—closer to a re-enterable closure than to a bare pointer.

That matters in MOOLLM because:

• Shared index — PROTOCOLS.yml, skills/INDEX.yml, and prior turns give the same symbol the same load-bearing role. Without that, a name is noise.
• Embodied triggers — A photo, a warning sign, a vocal cue can act like a K-line if the culture around it is shared: they snap attention to a lived situation, not a map coordinate.

Linguistic analogy (not Minsky’s wording, but useful): In some languages, am Bahnhof Zoo vs bare Bahnhof Zoo shifts from bare reference toward situated, insider scene—the way the 405 in Los Angeles is not “a highway” but that corridor of life. Same mechanism in spirit: name → thick situation, not only name → fact.

Optional cultural example: Nina Hagen’s Berlin Zoo (1982) is often read as performative activation—noise, animal theatre, West Berlin grit—not as zoology. The track works as a charged name for a subculture-tinted “zoo” of people and noise. A zoo sign about a lion spraying urine through bars is thematically on that wavelength: blunt, bodily, public—scene-first, not taxonomy-first.

Contrast: Frames (Minsky) describe the shape of a situation (slots, defaults). Drescher schemas (see schema-mechanism) describe learned causality (Context → Action → Result). Schemapedia (see schema, schemas/registry.yml) indexes interchange, causal, frames, activation, relational SQL/SQLite, and SoM—each family with its own delegate skills. Protocol symbols here describe which competence bundle to load. See knowledge-frames for the full map.

The 1979 memo mechanics

The memo (AI Memo 516; full OCR text in sources/aim-516-k-lines-1979-ocr.txt) gives the machinery, and every piece maps onto context engineering:

• K-node assignment + K-line attachment. A memorable event mints a K-node; its K-line makes an excitatory attachment to every currently active P-agent. A K-line is a stored attention mask over a P-pyramid: which nodes to activate and how much. Reactivation makes P "virtually hallucinate" the old partial state.
• The level-band principle (Minsky's "most important idea"). Attach only to an intermediate band: too low "would impose false perceptions and conceal the real details of the present problem"; too high "would make us hallucinate the present problem as already solved." Restore the middle of the pyramid — structure — leaving the top free for the current goal and the bottom free for current facts. Context-loading policy, 1979.
• Weak fringes = frame defaults. Connections at the band's edges are weak, so they lose cross-exclusion competition to anything the present situation asserts. Defaults that yield to evidence, implemented as low attention weights.
• Cross-exclusion. Agents cluster in mutually inhibiting groups — radio buttons. Force one on and it suppresses its rivals: cheap state reset, built-in short-term memory (persistence), and dispositions — a P-pyramid's repertory of pre-activated subsets, i.e. ways of seeing (the Necker cube flip). Conflict inside a group makes the whole group drop out and default upward: don't average contradictions, zoom out.
• K-recursion. New K-lines attach to currently active K-nodes, not raw P-agents: "new memories are composed mainly of ingredients from earlier memories." Vocabulary compounds — each named view becomes a component of the next.
• Crossbar via sparse coding. K-lines don't need full wiring; random small subsets of a shared line bundle suffice (Mooers zatocoding 1956, Willshaw 1969) — hashing and Bloom filters as neuroanatomy. Plus the Society answer: most agents never need to talk.

The P-pyramid — the agent-scoped hierarchical view a K-line masks over — has its own design doc: designs/P-PYRAMID.md.

Naming Convention

Why This Convention?

1. Grep-Friendly

# Find all uses of a protocol
grep -r "YAML-JAZZ" .

# Find all protocol definitions
grep -r "^[A-Z-]*:" PROTOCOLS.yml

# Find all mentions in docs
grep -rn "POSTEL\|PLAY-LEARN-LIFT" skills/

2. Vector-Search Friendly

The UPPER-CASE format creates distinct embeddings. When you search for "YAML-JAZZ", you get:

• The protocol definition
• Files that reference it
• Examples that demonstrate it

Not general YAML documentation.

3. Human-Scannable

In prose, protocols pop:

When parsing user input, apply POSTEL (charitable interpretation). If the command is ambiguous, use YAML-JAZZ to infer intent.

You can't miss them.

Protocol Layers

Kernel Protocols (Low-Level)

Defined in kernel/, fundamental to operation:

Skill Protocols (Mid-Level)

Defined in skills/, behavioral patterns:

Ad-Hoc Protocols (Emergent)

Mentioned in conversations, examples, documents:

# In a soul-chat document
We should apply KITCHEN-RULES here — no running,
always clean up, respect the fridge contents.

# In an adventure room
This room follows GRUE-SAFE protocol — magically lit.

These don't need formal definitions. Naming them makes them real.

Where Protocols Live

Central Registry

PROTOCOLS.yml — The master list:

YAML-JAZZ:
  meaning: "Comments carry meaning. Data is interpreted."
  location: kernel/constitution-core.md §3
  invoke_when: "Writing or reading YAML with semantic intent"

Embedded in Files

Protocols can be defined inline where they're used:

# In a room's ROOM.yml
protocols:
  DARK-ROOM:
    meaning: "No light source = grue danger"
    applies_when: "player enters without lit lamp"

Referenced in Prose

Just mention them:

This adventure follows REINCARNATION protocol — death returns you to start, inventory preserved.

The name IS the definition until someone formalizes it.

Protocol Lifecycle

Mentioned → Documented → Formalized → Skill
    ↓            ↓            ↓          ↓
  "Let's       Entry in     Directory   Full
   try          PROTOCOLS    in          PROTOTYPE
   FOOD-CHAIN"  .yml         skills/     + templates

Most protocols stay informal. Only crystallize when needed.

K-REFs: File Pointers as K-Lines

A K-REF is a K-line that points to a specific location in a file with optional metadata:

/path/to/file.md:42-58 # section_type | preview text

K-REF Format

PATH[:LINE[-END]][#ANCHOR][?SEARCH] # TYPE [LABEL] SEVERITY - DESCRIPTION
  EXCERPT or MASKED_VALUE

Components:

• PATH — Absolute file path (if no line number, refers to whole file)
• LINE — Line number (1-based), optionally with column :LINE:COL. Omit for whole file.
• END — Optional end line or column for ranges
• #ANCHOR — Jump to heading/section (like URL fragments)
• ?SEARCH — Find first match of pattern (like URL query)
• TYPE — What was found (uuid, secret, user_prompt, tool_call, etc.)
• LABEL — Optional redact label like [SSH_KEY]
• SEVERITY — Icon: 🔴 critical, 🟠 high, 🟡 medium, 🔵 low, ℹ️ info
• DESCRIPTION — Human-readable explanation
• EXCERPT — Surrounding context or masked value

Location Specifiers

Anchor Types

# Markdown headings (slugified)
/README.md#installation
/SKILL.md#k-ref-format

# YAML keys (dot-path)
/config.yml#server.port
/CARD.yml#methods.K-REF

# Code symbols (function/class)
/main.py#def:process_data
/main.py#class:AuditRunner

# HTML-style id
/page.html#footer

Search Patterns

# Literal string (default)
/file.py?def main

# Regex (prefix with ~)
/file.py?~def\s+\w+\(

# First match of pattern
/transcript.txt?user: 

# Combine with line context
/file.py?def main # function - Jump to main()

Examples

/path/to/config.yml # config - Whole file reference
/path/to/SKILL.md#k-refs # section - Jump to K-REFs section
/path/to/main.py?def audit # search - Find audit function
/path/transcript.txt:91:12-48 # uuid ℹ️
  fe18ce96-5200-4e15-acd3-190695de6295

/path/transcript.txt:7528:18-45 # private_key ([PRIVATE_KEY]) 🔴 - Private key header
  ********** ******* ********

/path/config.json:15 # password ([PASSWORD]) 🟠 - Database password
  ************

Images Work Too!

Cursor (and other orchestrators) can read absolute file paths directly — INCLUDING IMAGES!

/Users/me/Screenshots/error.png # screenshot - Analyze this error
/tmp/architecture-diagram.jpg # diagram - Explain this system
/path/to/chart.png # data - What trends do you see?

No line numbers for images (obviously). Just the absolute path.

This is powerful:

• Point to any image on disk
• Cursor reads and analyzes it
• Works with screenshots, diagrams, charts, photos
• Supports: jpeg, png, gif, webp

Security note: Cursor can read ANYWHERE on your disk, not just the workspace. Powerful, but be aware.

Sister Script Methodology

Problem: LLMs have limited context windows. Dumping entire files wastes tokens.

Solution: Sister scripts emit K-REFs (pointers) instead of full content:

# Sister script → K-REF → LLM flow
kref_flow:
  source: "cursor-mirror (sister script)"
  produces: "K-REFs (pointers)"
  consumer: "LLM reads only what it needs"

1. Scan — Script searches large data (transcripts, databases)
2. Emit — Output K-REFs pointing to relevant sections
3. Read — LLM selectively reads only the referenced ranges

This is reference by pointer, not by value — parsimonious context.

K-REF Output Modes

Example: cursor-mirror as Sister Script

# Emit K-REFs for secrets (LLM reads just the pointers)
cursor-mirror audit --patterns secrets --emit-kref

# Output:
# /path/file.txt:42:10-46 # openai_key ([OPENAI_KEY]) 🔴 - OpenAI API key
#   sk-proj-****...(48 chars)...****

# LLM can then selectively read context around line 42

Commands

Protocol Symbols

Dovetails With

• ../skill/ — Protocols can become skills
• ../../PROTOCOLS.yml — The central registry
• ../../kernel/ — Low-level protocols
• ../yaml-jazz/ — Protocol for semantic YAML
• ../postel/ — Protocol for charitable interpretation

The Power of Naming

"In the beginning was the Word."

When you name something, you can:

• Find it — grep, vector search, memory
• Invoke it — type the name, activate the tradition
• Compose it — combine named things into new things
• Teach it — names are handles for concepts

MOOLLM protocols are handles on ideas. The UPPER-CASE convention makes them impossible to miss, trivial to find.

"Name it, grep it, use it."