diff --git a/memory.md b/memory.md
index a40bf4c..d7914e2 100644
--- a/memory.md
+++ b/memory.md
@@ -293,6 +293,10 @@ Same scope as locked in `~/claude/ideas/rimlike/plan.md`. Realistic timeline 3
 - **Pattern recorded — ".import auto-generation via `godot --headless --import`".** Writing `*.png.import` companion files by hand with placeholder paths works but Godot rewrites them with proper UIDs on the first import scan. Reliable workflow: copy PNG into `art/`, write minimal .import (UID will be ignored), then run `/mnt/d/godot/Godot_v4.6.2-stable_win64_console.exe --headless --import` once to generate the .ctex binaries. Headless boot fails until the import scan runs. Same pattern from 2026-05-12 confirmed.
 - Delegation report (whole-day 05-15 polish sprint): **No delegation — all on Opus.** Sprite-atlas surveys, MCP runtime verification (camera positioning, dialog dismissal, stage forcing), and the iterative play→fix→play loop kept the same files hot in context; Haiku/Sonnet handoffs would have re-read every time. The `researcher` (Haiku) dispatch from earlier in the day for the job-system audit is the only delegated work.
 
+- **Workbenches rebuilt procedural after second atlas-misidentification.** The 2026-05-12 atlas-sprite pass for Carpenter/Smelter/Hearth/Millstone again hit the same trap as Bed: the picked tiles read as chest-of-drawers (Carpenter, FG_Interior 24,20), candle base (Smelter, FG_Marketplace 8,30), 2-burner stove (Hearth, FG_Interior 16,32), and cushion stack (Millstone, FG_Interior 17,40). User: "the carpenter is a chest of draws? no idea what the mill is, looks like a bucket". Survey of all 7 plausible source tilesets (Interior / Marketplace / Houses / Village / Fortress / Mines / Dark Castle / Extras) found a 2×2 medieval stone fireplace at FG_Interior (12,13)..(13,14) for Hearth but **2-wide tiles don't fit the current single-column sprite system**, and the other three had no clean replacement. Decision: go procedural for all four, following `CremationPyre._draw_pyre` as precedent. New methods `_draw_carpenter / _draw_smelter / _draw_hearth / _draw_millstone` in workbench.gd; `_VARIANT_SPRITES` table + `_build_sprite()` Sprite2D path removed; `_complete()` no longer touches a sprite child. Hearth is the only h=2 variant — draws stone surround + mantle + arched opening + log fire + flame teardrop extending into the tile north (Y-sort handles occlusion). Commit `c97ada8`.
+- **Pattern recorded — "atlas-misidentification trap, second occurrence — prefer procedural for named furniture".** Pixel-art tile atlases pack hundreds of squat 16×16 items at low resolution; without zoomed-in inspection it's easy to mistake a furniture silhouette for what its name implies. Both Bed (2026-05-12) and the four workbenches (2026-05-15) hit this failure mode. Going forward: for furniture whose silhouette is **definitional** (workbench-with-tools, bed-with-pillow, millstone-as-grindstone), **start with procedural draws** and only swap in an atlas tile after verifying via 16× zoom that the tile actually depicts the named object. Atlas-sourced sprites are still the right call for generic decorative variation (trees, walls, floors, crops with growth stages).
+- **Tree variety added via 3-season palette swap.** Copied `FG_Tree_Summer.png` + `FG_Tree_Fall.png` from the Grasslands tier-1 pack into `art/sprites/` (Spring was already there). 4 silhouettes × 3 palettes = 12 visual variants; `tree.gd` hash mixes silhouette and season independently so neighbouring tiles don't all share the same palette. Winter omitted — snowy trees would look out of place in the current biome. When a season-cycle system lands (future phase), the active texture can be swapped globally by season instead of per-tree. No growth stage / sapling system yet — bundle has no growth frames for trees, and the user's "regrow / auto-grow" ask is explicit future scope (deferred to a later phase that also covers planting designation). Commit `c97ada8`.
+
 ## External references
 
 - **Forgejo repo:** https://git.rdx4.com/megaproxy/rimlike (private)