Performance Targets

HEBBS defines hard latency budgets for every operation. These are contractual targets — regressions beyond these thresholds block releases.

Latency Targets

All targets are measured at p99 under load (1000 concurrent clients, 10M memories) unless otherwise noted.

Operation	p50 Target	p99 Target	Notes
remember	< 2 ms	< 5 ms	Excludes embedding (amortized via batching)
recall (similarity)	< 5 ms	< 10 ms	HNSW query at 10M memories, ef_search=100
recall (temporal)	< 2 ms	< 5 ms	B-tree range scan
recall (causal)	< 3 ms	< 8 ms	BFS traversal, max depth 5
recall (analogical)	< 8 ms	< 15 ms	Combined HNSW + cross-entity graph
get	< 1 ms	< 2 ms	Single-key RocksDB lookup
revise	< 3 ms	< 6 ms	Write new version + update indexes
forget (single)	< 2 ms	< 5 ms	Delete from all indexes
forget (entity)	varies	varies	Proportional to entity memory count
prime	< 5 ms	< 10 ms	Preload into OS page cache
embed (single)	< 1 ms	< 2 ms	Single embedding on CPU
embed (batch 32)	< 8 ms	< 12 ms	Amortized: ~0.3ms per item
health	< 0.5 ms	< 1 ms	No I/O

Throughput Targets

Metric	Target	Configuration
Remember (sustained)	> 3,000/sec	CPU, batch embedding
Remember (sustained)	> 10,000/sec	GPU (CUDA), batch embedding
Recall (concurrent)	> 5,000/sec	10M memories, similarity
Subscribe pushes	> 1,000/sec	Per subscription stream

Embedding Latency Budget

Embedding is the most variable cost. The targets assume batch amortization:

Mode	Per-Item Latency	Batch Size
CPU (single)	1–2 ms	1
CPU (batched)	0.25–0.4 ms	32
CoreML (Apple Silicon)	0.1–0.3 ms	32
CUDA (NVIDIA)	0.05–0.15 ms	32

Reflection Pipeline

Reflection runs in the background and is not subject to hot-path latency budgets. Targets are for overall pipeline completion:

Metric	Target
Cluster stage	< 500 ms per 1000 memories
Propose stage	Depends on LLM provider latency
Validate stage	< 100 ms per insight candidate
Consolidate stage	< 50 ms per validated insight
Full cycle (1000 memories)	< 30 seconds (excluding LLM)

Measurement

All targets are measured using:

• Criterion benchmarks in the benches/ directory for micro-benchmarks
• hebbs-bench CLI tool for end-to-end load testing
• Prometheus histograms for production monitoring

A regression > 10% on any p99 target blocks the merge. See Running Benchmarks for how to measure.