diff --git a/src/assemblers/caches/README.md b/src/assemblers/caches/README.md
new file mode 100644
index 0000000..82570d6
--- /dev/null
+++ b/src/assemblers/caches/README.md
@@ -0,0 +1,16 @@
+# src/assemblers/caches/
+
+Pre-allocated per-element scratch and global COO/CSC caches reused by
+every assembly strategy. Keeping them in one place makes it easy to see
+what is allocated up-front and what is consumed by the hot loops.
+
+## Files
+
+- `geometry_cache.jl` — `GeometryCache`. Per-element shape values, derivatives, and Jacobian data sized for one quadrature rule. Parametric on storage so the same struct works heap-owned (element-based assemblers) or as a view into batched SoA storage (DOF-based assembler).
+- `element_cache.jl` — `ElementCache`. Assembly-time per-element scratch (local stiffness, force, dof index buffers).
+- `material_cache.jl` — `AssemblyMaterialWorkspace` plus the `create_material_cache` helper. Holds the per-IP material state buffers (stress, tangent, history) used by the kernels.
+- `coo_cache.jl` — `COOCache`, the global I/J/V triplet workspace used by the element-based assembler.
+
+The hot-path invariant for all of these is zero allocation per element
+on a steady-state assemble. The regression test for that is
+`test/assemblers/test_dof_based_zero_alloc.jl`.