diff --git a/stats_synthetic_bench_test.go b/stats_synthetic_bench_test.go
new file mode 100644
index 00000000..58571829
--- /dev/null
+++ b/stats_synthetic_bench_test.go
@@ -0,0 +1,183 @@
+package stats
+
+import "testing"
+
+// syntheticTrace simulates the call pattern of a request-serving service
+// that holds a long-lived root Scope but re-derives nested scopes and
+// tagged metric handles afresh, inline, on every request - rather than
+// caching subScope/Counter/Gauge/Timer references itself. This is the exact
+// pattern that made gostats' own joinScopes/MergeTags/Serialize path show up
+// as a large fraction of allocations under profiling: every call below is,
+// from the caller's perspective, "the same metric" as the last call, but
+// prior to caching gostats had no way to know that without redoing the full
+// key computation from scratch each time.
+//
+// NewPerInstanceCounter is deliberately excluded here (see
+// BenchmarkSyntheticTracePerInstance) - it isn't memoized (see stats.go), so
+// mixing it in would stop this benchmark from showing the true 0-allocs/op
+// steady state of the calls that are.
+func syntheticTrace(root Scope) {
+	graph := root.ScopeWithTags("graph", map[string]string{"graph_name": "supply_estimate"})
+	node := graph.ScopeWithTags("node", map[string]string{"source_node": "waypoint_bonus"})
+
+	regionTags := map[string]string{"region_code": "us-east-1", "ride_type": "standard"}
+	node.NewCounterWithTags("attempt", regionTags).Inc()
+	node.NewCounterWithTags("success", regionTags).Inc()
+	node.NewGaugeWithTags("queue_depth", map[string]string{"region_code": "us-east-1"}).Set(5)
+	node.NewTimerWithTags("latency", map[string]string{"region_code": "us-east-1"}).AddValue(1500)
+
+	fee := root.ScopeWithTags("fee", map[string]string{"fee_type": "toll"}).
+		ScopeWithTags("region", map[string]string{"region_code": "us-east-1"})
+	fee.NewCounterWithTags("charged", map[string]string{"toll_period": "peak"}).Inc()
+	fee.NewCounterWithTags("waived", map[string]string{"toll_period": "peak"}).Inc()
+}
+
+// BenchmarkSyntheticTrace runs syntheticTrace repeatedly against a single,
+// long-lived root Scope - the steady-state condition once a service has
+// warmed up and every (name, tags) combination below has already been seen
+// at least once. Compare -benchmem output against the pre-memoization
+// revision to see the allocation delta from caching Scope/Counter/Gauge/
+// Timer lookups (see the "Memoize tag-based Scope/Counter/Gauge/Timer
+// lookups" commit).
+func BenchmarkSyntheticTrace(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	root := store.Scope("app")
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		syntheticTrace(root)
+	}
+}
+
+// BenchmarkSyntheticTraceParallel is the same trace under concurrent load
+// (multiple goroutines sharing the same root Scope), which is the realistic
+// shape for a request-serving service.
+func BenchmarkSyntheticTraceParallel(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	root := store.Scope("app")
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			syntheticTrace(root)
+		}
+	})
+}
+
+// pooledTags holds one map[string]string per distinct tag set used by
+// syntheticTracePooled, built once and reused across every call instead of
+// allocated fresh inline. gostats never mutates a tags map handed to it, so
+// sharing these across goroutines (BenchmarkSyntheticTracePooledParallel) is
+// safe.
+type pooledTags struct {
+	graph, node, region, queueDepth, latency, fee, feeRegion, toll, dispatch map[string]string
+}
+
+func newPooledTags() *pooledTags {
+	return &pooledTags{
+		graph:      map[string]string{"graph_name": "supply_estimate"},
+		node:       map[string]string{"source_node": "waypoint_bonus"},
+		region:     map[string]string{"region_code": "us-east-1", "ride_type": "standard"},
+		queueDepth: map[string]string{"region_code": "us-east-1"},
+		latency:    map[string]string{"region_code": "us-east-1"},
+		fee:        map[string]string{"fee_type": "toll"},
+		feeRegion:  map[string]string{"region_code": "us-east-1"},
+		toll:       map[string]string{"toll_period": "peak"},
+		dispatch:   map[string]string{"region_code": "us-east-1"},
+	}
+}
+
+// syntheticTracePooled is syntheticTrace with every tags map built once (via
+// newPooledTags) and reused across calls, as a caller might do by holding a
+// package-level or per-instance pool of tag maps instead of building a fresh
+// map[string]string literal on every call. This removes the caller-side map
+// allocations from the measurement, isolating gostats' own internal
+// allocations (the joinScopes/MergeTags/Serialize path and the scopeCache
+// bookkeeping around it) so the -benchmem delta between the pre- and
+// post-memoization revisions reflects gostats' contribution alone.
+//
+// Like syntheticTrace, NewPerInstanceCounter is excluded - see
+// BenchmarkSyntheticTracePerInstancePooled.
+func syntheticTracePooled(root Scope, p *pooledTags) {
+	graph := root.ScopeWithTags("graph", p.graph)
+	node := graph.ScopeWithTags("node", p.node)
+
+	node.NewCounterWithTags("attempt", p.region).Inc()
+	node.NewCounterWithTags("success", p.region).Inc()
+	node.NewGaugeWithTags("queue_depth", p.queueDepth).Set(5)
+	node.NewTimerWithTags("latency", p.latency).AddValue(1500)
+
+	fee := root.ScopeWithTags("fee", p.fee).ScopeWithTags("region", p.feeRegion)
+	fee.NewCounterWithTags("charged", p.toll).Inc()
+	fee.NewCounterWithTags("waived", p.toll).Inc()
+}
+
+// BenchmarkSyntheticTracePooled is BenchmarkSyntheticTrace with pooled tag
+// maps (see syntheticTracePooled) - the allocs/op and B/op it reports are
+// gostats' own internal cost, with the caller-side map-literal allocations
+// that BenchmarkSyntheticTrace also pays for removed from the measurement.
+func BenchmarkSyntheticTracePooled(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	root := store.Scope("app")
+	tags := newPooledTags()
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		syntheticTracePooled(root, tags)
+	}
+}
+
+// BenchmarkSyntheticTracePooledParallel is BenchmarkSyntheticTracePooled
+// under concurrent load, sharing both the root Scope and the pooled tags
+// across goroutines.
+func BenchmarkSyntheticTracePooledParallel(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	root := store.Scope("app")
+	tags := newPooledTags()
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			syntheticTracePooled(root, tags)
+		}
+	})
+}
+
+// BenchmarkSyntheticTracePerInstance measures NewPerInstanceCounter alone,
+// with a fresh tags map[string]string literal built inline each call (as
+// syntheticTrace does for its other calls). NewPerInstanceCounter is not
+// memoized by scopeCache (see stats.go), so this should show no allocation
+// improvement over the pre-memoization revision - it's a control, not a
+// regression.
+func BenchmarkSyntheticTracePerInstance(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	node := store.Scope("app").ScopeWithTags("graph", map[string]string{"graph_name": "supply_estimate"}).
+		ScopeWithTags("node", map[string]string{"source_node": "waypoint_bonus"})
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		node.NewPerInstanceCounter("dispatch", map[string]string{"region_code": "us-east-1"}).Inc()
+	}
+}
+
+// BenchmarkSyntheticTracePerInstancePooled is
+// BenchmarkSyntheticTracePerInstance with the tags map pooled (see
+// pooledTags), isolating NewPerInstanceCounter's own internal allocations
+// from the caller-side map-literal cost.
+func BenchmarkSyntheticTracePerInstancePooled(b *testing.B) {
+	store := NewStore(NewNullSink(), false)
+	node := store.Scope("app").ScopeWithTags("graph", map[string]string{"graph_name": "supply_estimate"}).
+		ScopeWithTags("node", map[string]string{"source_node": "waypoint_bonus"})
+	tags := newPooledTags()
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		node.NewPerInstanceCounter("dispatch", tags.dispatch).Inc()
+	}
+}