diff --git a/mandel.s b/mandel.s
index 097b700..467245d 100644
--- a/mandel.s
+++ b/mandel.s
@@ -21,12 +21,16 @@ count_pixels = $a3 ; u8
 total_ms     = $a4 ; float48
 total_pixels = $aa ; float48
 
-temp         = $b0 ; u16
-pixel_ptr    = $b2 ; u16
-pixel_color  = $b4 ; u8
-pixel_mask   = $b5 ; u8
-pixel_shift  = $b6 ; u8
-pixel_offset = $b7 ; u8
+z_buffer_active = $b0 ; boolean: 1 if we triggered the lake, 0 if not
+z_buffer_start  = $b1 ; u8: index into z_buffer
+z_buffer_end    = $b2 ; u8: index into z_buffer
+temp            = $b4 ; u16
+
+pixel_ptr       = $b6 ; u16
+pixel_color     = $b8 ; u8
+pixel_mask      = $b9 ; u8
+pixel_shift     = $ba ; u8
+pixel_offset    = $bb ; u8
 
 
 ; FP registers in zero page
@@ -186,6 +190,15 @@ color_map:
 
 .code
 
+z_buffer_len = 16
+z_buffer_mask = z_buffer_len - 1
+z_buffer:
+    ; the last N zx/zy values
+    .repeat z_buffer_len
+        .word 0
+        .word 0
+    .endrepeat
+
 .export start
 
 ; 2 + 9 * byte cycles
@@ -462,12 +475,14 @@ initloop:
     sta zx - 1,x
     dex
     bne initloop
+    sta z_buffer_start
+    sta z_buffer_len
 
 loop:
     ; iter++ & max-iters break
     inc iter
     bne keep_going
-    rts
+    jmp exit_path
 keep_going:
 
     .macro quick_exit arg, max
@@ -484,7 +499,7 @@ keep_going:
     positive:
         cmp #((max) << 4)
         bmi all_done ; 'less than'
-        rts
+        jmp exit_path
 
     negative:
         cmp #(256 - ((max) << 4))
@@ -492,7 +507,7 @@ keep_going:
         bpl all_done    ; 'greater than'
 
     nope_out:
-        rts
+        jmp exit_path
     
     first_equal:
         lda arg
@@ -527,9 +542,90 @@ keep_going:
 
     ; if may be in the lake, look for looping output with a small buffer
     ; as an optimization vs running to max iters
+    lda z_buffer_active
+    beq skip_z_buffer
+
+    ldx z_buffer_start
+    cpx z_buffer_end
+    beq z_nothing_to_read
+
+z_buffer_loop:
+    .macro z_compare arg
+        .local compare_no_match
+        lda z_buffer,x
+        inx
+        cmp arg
+        bne compare_no_match
+        iny
+    compare_no_match:
+    .endmacro
+    .macro z_advance
+        .local skip_reset_x
+        cpx #(z_buffer_len * 4)
+        bmi skip_reset_x
+        ldx #0
+    skip_reset_x:
+    .endmacro
+    .macro z_store arg
+        lda arg
+        sta z_buffer,x
+        inx
+    .endmacro
+
+    ; Compare the previously stored z values
+    ldy #0
+    z_compare zx
+    z_compare zx + 1
+    z_compare zy
+    z_compare zy + 1
+
+    cpy #4
+    bne z_no_matches
+    jmp z_exit
+
+z_no_matches:
+    z_advance
+
+    cpx z_buffer_end
+    bne z_buffer_loop
+
+z_nothing_to_read:
+
+    ; Store and expand
+    z_store zx
+    z_store zx + 1
+    z_store zy
+    z_store zy + 1
+    z_advance
+    stx z_buffer_end
+
+    ; Increment the start roller if necessary (limit size)
+    lda iter
+    cmp #(z_buffer_len * 4)
+    bmi skip_inc_start
+    lda z_buffer_start
+    clc
+    adc #4
+    tax
+    z_advance
+    stx z_buffer_start
+skip_inc_start:
+
+skip_z_buffer:
+
     jmp loop
 
-peace_out:
+z_exit:
+    lda #0
+    sta iter
+
+exit_path:
+    ldx #0
+    lda iter
+    bne next
+    inx
+next:
+    stx z_buffer_active
     rts
 
 .endproc