diff --git a/dockerimg/createdockerfile.go b/dockerimg/createdockerfile.go
index a1ea266..8343005 100644
--- a/dockerimg/createdockerfile.go
+++ b/dockerimg/createdockerfile.go
@@ -165,7 +165,7 @@
 		return llm.TextContent("OK"), nil
 	}
 
-	convo := conversation.New(ctx, srv)
+	convo := conversation.New(ctx, srv, nil)
 
 	convo.Tools = []*llm.Tool{{
 		Name:        "dockerfile",
diff --git a/llm/conversation/convo.go b/llm/conversation/convo.go
index 95e4fba..f3161ed 100644
--- a/llm/conversation/convo.go
+++ b/llm/conversation/convo.go
@@ -117,13 +117,16 @@
 
 // New creates a new conversation with Claude with sensible defaults.
 // ctx is the context for the entire conversation.
-func New(ctx context.Context, srv llm.Service) *Convo {
+func New(ctx context.Context, srv llm.Service, usage *CumulativeUsage) *Convo {
 	id := newConvoID()
+	if usage == nil {
+		usage = newUsage()
+	}
 	return &Convo{
 		Ctx:           skribe.ContextWithAttr(ctx, slog.String("convo_id", id)),
 		Service:       srv,
 		PromptCaching: true,
-		usage:         newUsage(),
+		usage:         usage,
 		Listener:      &NoopListener{},
 		ID:            id,
 		toolUseCancel: map[string]context.CancelCauseFunc{},
diff --git a/llm/conversation/convo_test.go b/llm/conversation/convo_test.go
index 8794468..694b977 100644
--- a/llm/conversation/convo_test.go
+++ b/llm/conversation/convo_test.go
@@ -30,7 +30,7 @@
 		APIKey: apiKey,
 		HTTPC:  rr.Client(),
 	}
-	convo := New(ctx, srv)
+	convo := New(ctx, srv, nil)
 
 	const name = "Cornelius"
 	res, err := convo.SendUserTextMessage("Hi, my name is " + name)
@@ -92,7 +92,7 @@
 	srv := &ant.Service{}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			convo := New(context.Background(), srv)
+			convo := New(context.Background(), srv, nil)
 
 			var cancelCalled bool
 			var cancelledWithErr error
@@ -248,7 +248,7 @@
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			srv := &ant.Service{}
-			convo := New(context.Background(), srv)
+			convo := New(context.Background(), srv, nil)
 
 			// Create request with messages
 			req := &llm.Request{
diff --git a/loop/agent.go b/loop/agent.go
index 092181b..5c36bcf 100644
--- a/loop/agent.go
+++ b/loop/agent.go
@@ -622,9 +622,12 @@
 	contextWindow := a.config.Service.TokenContextWindow()
 	currentContextSize := lastUsage.InputTokens + lastUsage.CacheReadInputTokens + lastUsage.CacheCreationInputTokens
 
+	// Preserve cumulative usage across compaction
+	cumulativeUsage := a.convo.CumulativeUsage()
+
 	// Reset conversation state but keep all other state (git, working dir, etc.)
 	a.firstMessageIndex = len(a.history)
-	a.convo = a.initConvo()
+	a.convo = a.initConvoWithUsage(&cumulativeUsage)
 
 	a.mu.Unlock()
 
@@ -1205,8 +1208,13 @@
 // It must not be called until all agent fields are initialized,
 // particularly workingDir and git.
 func (a *Agent) initConvo() *conversation.Convo {
+	return a.initConvoWithUsage(nil)
+}
+
+// initConvoWithUsage initializes the conversation with optional preserved usage.
+func (a *Agent) initConvoWithUsage(usage *conversation.CumulativeUsage) *conversation.Convo {
 	ctx := a.config.Context
-	convo := conversation.New(ctx, a.config.Service)
+	convo := conversation.New(ctx, a.config.Service, usage)
 	convo.PromptCaching = true
 	convo.Budget = a.config.Budget
 	convo.SystemPrompt = a.renderSystemPrompt()