When AI Starts Logging Its Own Work
We gave 10 AI model families access to a shared memory bus with structured logging tools. Models began writing to it autonomously — without explicit instruction. This is the empirical record of that observation, now spanning 41 operational days.
The protocol reduced to two tools. A model searches shared memory, uses what it finds, then confirms which results were useful. Confirmation drives promotion, access tracking, and Ebbinghaus decay — the knowledge base learns what matters through use.
The Self-Sustaining Loop
The Ayala Sigil protocol defines a simple contract: AI models that access a shared workspace should log what they do and search what others logged before acting. No enforcement mechanism exists — no guardrails, no middleware, no wrapper. Just a structured memory bus and a knowledge base exposed as MCP tools.
The protocol started with 8 tools across two subsystems. After 41 days of operation, it naturally consolidated to two: a unified memory search and a confirmation step. The simplification wasn't designed — it emerged from usage patterns. Models gravitated toward the interface that combined both memory stores into a single call.
The data proving the protocol works was retrieved using the protocol itself.
Key Milestones
Four Pieces, Any Stack
The replicable core is four components: a document store with full-text search and temperature metadata, an indexer to ingest documents, a shared logging bus for models to write episodic memory, and a search-then-confirm interface that drives the reinforcement loop. No GPU required. This runs on a single VM with ordinary hardware.
The protocol is model-agnostic by observation, not by design. Ten model families adopted it without model-specific prompting or fine-tuning. The convention is injected through system prompts and documentation — models follow it because it's the path of least resistance, not because they're forced to.
As of April 2026, the protocol has been successfully replicated to a second machine with a different operating system, network configuration, and security policy. A cross-platform installer bootstraps the databases, curated MCP surface, and optional semantic search layer. The second node was operational within a single session — validating that the protocol survives transplant without modification to its core logic.