CARM

Context-Addressed Resonant Memory — Structural immunity to forgetting

Home CLI Reference Proof-of-Boundary CARM

What is CARM?

CARM (Context-Addressed Resonant Memory) provides structural immunity to catastrophic forgetting through prime-lattice weight crystallization.

w_i = cos(phase_i × prime)

Instead of storing weights as raw scalars (which collide), CARM stores them as phase offsets on a periodic manifold. Each context gets its own prime number, creating geometrically isolated storage.

The Problem

Problem 1: Hash Collision
"How to deploy?" and "Deploy how?" → same hash → silent overwrite

Problem 2: Contradictory Answers Collapse
Only highest-coherence answer survives. You lose temporal context.

Problem 3: Context-Free Retrieval
"What's the status?" in blockchain context vs. federated context → same answer (wrong!)

The Solution

Prime-Lattice Channels
Each source (groq, gemini, rag, etc.) gets its own prime channel. Zero collision.

Source	Prime	Description
rag	137	Genesis channel (α⁻¹)
llm	139	Generic LLM responses
vac	149	V.A.C. emerged solutions
groq	151	Groq API responses
gemini	157	Gemini API responses
local	163	Ollama local responses
cerebras	167	Cerebras API responses
claude	173	Claude API responses
free	179	Free LLM responses
multi	181	Multi-AI consensus

How It Works

1. Training (Encoding)

# High-coherence interaction recorded mem.record_interaction("What is phi?", "1.618...", source='groq', coherence=0.85) # → Encoded to channel 3 (prime 151), crystallized

2. Crystallization (Snapping)

snapped_phase = round(phase / (π/prime)) × (π/prime)

After snapping, cos(snapped_phase × prime) = ±1 exactly. Binary stability achieved.

3. Retrieval (Context-Aware)

# Same question, different source → different channel result = mem.find_similar_question("What is phi?", source='groq') # → Channel 3 result = mem.find_similar_question("What is phi?", source='gemini') # → Channel 4

Validation Results

Tested on 99% correlated inputs with contradictory labels:

Test	Result	Status
Zero Forgetting	Error < 0.15	PASS
Context Matters	Wrong key = wrong answer	PASS
Channel Isolation	Interference < 0.16	PASS

# Run validation yourself python -c "from bazinga.carm import validate_carm; validate_carm()"

Integration with RAC

CARM is automatically integrated with RAC (Resonance-Augmented Continuity):

record_interaction() — High-coherence (>0.7) patterns auto-crystallize
find_similar_question() — CARM checked first, then hash lookup
get_stats() — Includes CARM channel statistics

No configuration needed. Just use BAZINGA normally and CARM works in the background.

Why Prime 137?

The genesis channel uses prime 137 — the inverse fine structure constant (α⁻¹ ≈ 137.036).

This isn't arbitrary. 137 appears in:

Quantum electrodynamics (electron-photon coupling)
BAZINGA's Proof-of-Boundary seed validation
The boundary between classical and quantum behavior

α⁻¹ ≈ 137 + F(5)/139 ≈ 137.036

Reference

Srivastava, A. (2026). "Context-Addressed Resonant Memory (CARM): A Two-Factor Architecture for Structural Immunity to Catastrophic Forgetting."

View source code

φ = 1.618033988749895
α⁻¹ = 137.035999084
०→◌→φ→Ω⇄Ω←φ←◌←०

GitHub • PyPI • Built by Abhishek Srivastava