my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	"""
#2	BEAM Scale Benchmark — demonstrates recall stays flat as corpus grows.
#3	Run: PYTHONPATH=. python tests/benchmark_beam_scale.py
#4	"""
#5
#6	import time
#7	import tempfile
#8	from pathlib import Path
#9
#10	from mnemosyne.core.beam import BeamMemory
#11
#12
#13	def benchmark():
#14	with tempfile.TemporaryDirectory() as tmpdir:
#15	db_path = Path(tmpdir) / "scale.db"
#16	beam = BeamMemory(session_id="scale", db_path=db_path)
#17
#18	sizes = [100, 500, 1000, 2000]
#19	print("🔍 BEAM Recall Latency vs Corpus Size")
#20	print("-" * 50)
#21
#22	cumulative = 0
#23	for size in sizes:
#24	# Insert batch
#25	print(f"Inserting batch to reach {size} episodic memories...")
#26	t0 = time.time()
#27	batch = size - cumulative
#28	for i in range(batch):
#29	beam.consolidate_to_episodic(
#30	summary=f"Scale test item {cumulative + i}: concept {i % 100} in domain {(cumulative + i) % 10}",
#31	source_wm_ids=[f"s{cumulative + i}"],
#32	importance=0.5
#33	)
#34	insert_sec = time.time() - t0
#35	cumulative = size
#36	print(f" Batch insert ({batch} items): {insert_sec:.1f}s")
#37
#38	# Benchmark recall
#39	queries = ["concept 42", "domain 7", "nonexistent xyz"]
#40	for q in queries:
#41	times = []
#42	for _ in range(10):
#43	t0 = time.time()
#44	results = beam.recall(q, top_k=5)
#45	times.append((time.time() - t0) * 1000)
#46	avg = sum(times) / len(times)
#47	p95 = sorted(times)[int(len(times) * 0.95)]
#48	print(f" Corpus={size:4d} \| Query='{q[:20]:<20}' \| {avg:.2f}ms avg \| {p95:.2f}ms p95")
#49
#50	ep_stats = beam.get_episodic_stats()
#51	print(f"\n📊 Final episodic memory: {ep_stats['total']} items \| vectors: {ep_stats['vectors']}")
#52
#53
#54	if __name__ == "__main__":
#55	benchmark()
#56

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public