diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
new file mode 100644
index 00000000..252b38f6
--- /dev/null
+++ b/.github/workflows/benchmark.yml
@@ -0,0 +1,70 @@
+name: benchmarks
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+  schedule:
+    - cron: '23 3 * * *'   # nightly ~03:23 UTC
+  workflow_dispatch:
+
+jobs:
+  bench-smoke:
+    name: Benchmark smoke (compile + run 1x)
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: aide
+    steps:
+      - uses: actions/checkout@v7
+
+      - name: Setup Go
+        uses: actions/setup-go@v7
+        with:
+          go-version-file: aide/go.mod
+          cache-dependency-path: aide/go.sum
+
+      - name: Download dependencies
+        run: go mod download
+
+      # 1x each: proves every benchmark compiles and runs, cheap. This is the
+      # PR gate. NOTE: until the benchmark _test.go files land (dedicated PR),
+      # this passes vacuously (go test -bench=. with no benchmarks exits 0).
+      - name: Benchmark smoke test
+        run: make bench-smoke
+
+  bench-measure:
+    name: Benchmark measurement
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: aide
+    steps:
+      - uses: actions/checkout@v7
+
+      - name: Setup Go
+        uses: actions/setup-go@v7
+        with:
+          go-version-file: aide/go.mod
+          cache-dependency-path: aide/go.sum
+
+      - name: Download dependencies
+        run: go mod download
+
+      - name: Run benchmarks
+        run: |
+          go test -bench=. -benchmem -benchtime=200ms -count=3 \
+            ./pkg/store/... ./pkg/survey/... ./pkg/grammar/... ./pkg/code/... ./pkg/memory/... \
+            > bench-results.txt 2>&1
+          tail -5 bench-results.txt
+
+      # path is relative to GITHUB_WORKSPACE (repo root), not working-directory,
+      # so the file written inside aide/ is aide/bench-results.txt here.
+      - name: Upload benchmark results
+        uses: actions/upload-artifact@v7
+        with:
+          name: go-benchmarks
+          path: aide/bench-results.txt
+          retention-days: 14
+          if-no-files-found: error
diff --git a/Makefile b/Makefile
index c3c3c5f7..3fbb13c1 100644
--- a/Makefile
+++ b/Makefile
@@ -7,7 +7,7 @@
 #   make release-push VERSION=1.2.0  Bump to specific, commit, tag, and push
 #   make hooks                       Install git hooks (lefthook)
 
-.PHONY: release release-push build build-pprof build-web test test-ts test-go lint check-version check-release-needed hooks hooks-check
+.PHONY: release release-push build build-pprof build-web test test-ts test-go lint check-version check-release-needed hooks hooks-check bench bench-smoke
 
 VERSION_FILES = package.json .claude-plugin/plugin.json .claude-plugin/marketplace.json .codex-plugin/plugin.json packages/opencode-plugin/package.json $(wildcard packages/aide-binary-*/package.json)
 
@@ -127,6 +127,12 @@ test: hooks-check test-go test-ts
 test-go:
 	$(MAKE) -C aide test
 
+bench:
+	$(MAKE) -C aide bench
+
+bench-smoke:
+	$(MAKE) -C aide bench-smoke
+
 test-ts:
 	bunx vitest run --exclude='tests/memory-capture.test.ts' --exclude='dist/**'
 
diff --git a/aide/Makefile b/aide/Makefile
index 7009e034..ee1f552a 100644
--- a/aide/Makefile
+++ b/aide/Makefile
@@ -1,6 +1,6 @@
 # aide Makefile
 
-.PHONY: all build test clean lint fmt install install-plugin proto calibrate-tokens
+.PHONY: all build test clean lint fmt install install-plugin proto calibrate-tokens bench bench-smoke bench-profile
 
 # Binary name
 BINARY=aide
@@ -84,6 +84,38 @@ test-coverage:
 	$(GOCMD) tool cover -html=coverage.out -o coverage.html
 	@echo "Coverage report: coverage.html"
 
+# Full benchmark run with allocations, 3 runs for stability. Packages are the
+# hot-path areas that have (or will have) _bench_test.go files: store (bbolt +
+# bleve code index and memory), survey (repo analysis), grammar (project scan),
+# code (parse/tokenize), memory (scoring).
+bench:
+	@echo "Running Go benchmarks..."
+	$(GOTEST) -bench=. -benchmem -benchtime=200ms -count=3 \
+	  ./pkg/store/... ./pkg/survey/... ./pkg/grammar/... ./pkg/code/... ./pkg/memory/...
+
+# Benchmark run with CPU + heap profiles written to pprof-out/. One go test
+# invocation per package, since the profiler disallows writing profiles from
+# multiple packages in a single run. Analyze with:
+#   go tool pprof pprof-out/store.cpu.out
+bench-profile:
+	@mkdir -p pprof-out
+	@for pkg in store survey grammar code memory; do \
+		echo "Profiling pkg/$$pkg ..."; \
+		$(GOTEST) -bench=. -benchmem -benchtime=1s -count=1 \
+			-cpuprofile=pprof-out/$$pkg.cpu.out \
+			-memprofile=pprof-out/$$pkg.mem.out \
+			./pkg/$$pkg/... || exit 1; \
+	done
+	@ls -la pprof-out
+
+# Fast check that every benchmark compiles and executes without error (one
+# iteration each). Used in CI on every PR to catch broken benchmarks cheaply
+# without paying for a meaningful measurement run.
+bench-smoke:
+	@echo "Running Go benchmark smoke test (1 iteration each)..."
+	$(GOTEST) -run XXXNOTHING -bench=. -benchtime=1x -count=1 \
+	  ./pkg/store/... ./pkg/survey/... ./pkg/grammar/... ./pkg/code/... ./pkg/memory/...
+
 lint:
 	@echo "Running linter..."
 	@if command -v $(GOLINT) > /dev/null; then \