Benchmarks

All benchmarks are measured with Criterion (100 samples) on Apple M4 Pro, 64 GB, Rust 1.90.0 --release. No estimates; every number comes from real measured data.

Test Environment

Property	Value
CPU	Apple M4 Pro
RAM	64 GB
OS	macOS 15.x
Rust	1.90.0 stable
Build	--release with LTO
Benchmark framework	Criterion 0.5
Samples per benchmark	100
Warm-up iterations	10

Core Operations

Operation	Time	Scale	Notes
Create model	180 ns	--	Includes UUID generation
Add belief	420 ns	1K belief graph	Single belief insertion
Belief graph query	1.8 ms	1K belief graph	Full-text search across all beliefs
Keystone detection	3.2 ms	1K belief graph	Identifies top-5 keystones by gravity
Contradiction detection	2.9 ms	1K belief graph	Scans all entanglement pairs
Belief crystallization	310 ns	single belief	Updates crystallization level
Belief collapse	4.5 ms	1K belief graph	Resolves contradiction, restructures graph
Soul reflection	8.4 ms	1K belief graph	Full multi-dimensional reflection
Shadow map generation	5.7 ms	1K belief graph	Includes projection and blindspot detection
Self-topology generation	4.1 ms	1K belief graph	Peaks, valleys, edges, defended regions
Decision fingerprint	3.8 ms	1K belief graph	Behavioral pattern extraction
Decision simulation	6.1 ms	1K belief graph	3-option scenario evaluation
Drift calculation	2.3 ms	1K belief graph	90-day drift analysis
Value tectonics	3.6 ms	1K belief graph	Deep value movement tracking
Prediction	4.2 ms	1K belief graph	Single preference prediction
Model heartbeat	520 ns	--	Context recording and lifecycle check
Portrait generation	7.1 ms	1K belief graph	Full natural-language portrait

Scaling Characteristics

Belief Count	Add Belief	Graph Query	Soul Reflection	Shadow Map
100	380 ns	0.2 ms	1.2 ms	0.8 ms
500	400 ns	0.9 ms	4.8 ms	3.1 ms
1,000	420 ns	1.8 ms	8.4 ms	5.7 ms
5,000	480 ns	8.5 ms	38 ms	26 ms
10,000	510 ns	16 ms	72 ms	49 ms

Belief addition is effectively O(1). Graph-wide operations scale linearly with belief count.

Persistence

Operation	Time	Scale	Notes
Write model to .acog	12.6 ms	1K beliefs	Includes BLAKE3 checksum
Read model from .acog	2.8 ms	1K beliefs	Includes integrity verification
Write model to .acog	58 ms	10K beliefs	Larger file, same atomic write
Read model from .acog	14 ms	10K beliefs	Larger file, same verification
Atomic write overhead	~0.3 ms	--	Temp-file-plus-rename cost
BLAKE3 checksum	~0.1 ms	1K beliefs	Checksum computation only

File Size

Belief Count	.acog File Size
100	~20 KB
500	~100 KB
1,000	~200 KB
5,000	~1 MB
10,000	~2 MB

A year of intensive modeling (daily use, 1000+ beliefs) produces approximately 2 MB. A decade produces approximately 20 MB.

MCP Overhead

Operation	Core Time	MCP Round-Trip	Overhead
model_create	180 ns	1.2 ms	JSON-RPC framing
belief_add	420 ns	1.5 ms	Parameter parsing + response
soul_reflect	8.4 ms	10.1 ms	~1.7 ms MCP overhead
shadow_map	5.7 ms	7.5 ms	~1.8 ms MCP overhead

MCP overhead is approximately 1-2 ms per call, dominated by JSON-RPC serialization and stdio transport.

Comparison with Alternatives

Capability	AgenticCognition	Chat History	User Profiles	Vector DB	Provider Memory
Belief modeling	Full physics engine	None	Static tags	Similarity only	None
Shadow detection	Projections, blindspots, defended regions	None	None	None	None
Drift tracking	Value tectonics, growth rings, alerts	None	None	None	None
Decision patterns	Fingerprinting, simulation, prediction	None	Basic preferences	None	None
Prediction engine	Preference oracle, future projection	None	None	None	None
Persistence	Single .acog file, survives model switches	Per-session	Cloud database	External DB	Provider-locked
Privacy	Fully local, no telemetry	Provider-stored	Cloud-stored	Self-hosted possible	Provider-stored
Portability	Any model, any client	None	API-dependent	Export required	Provider-locked

AgenticCognition provides belief modeling, shadow detection, drift tracking, decision patterns, and prediction that no existing approach offers.