diff --git a/src/tensor.cpp b/src/tensor.cpp
index 257c396c3..9588bf224 100644
--- a/src/tensor.cpp
+++ b/src/tensor.cpp
@@ -476,10 +476,10 @@ static inline map<TensorVar, TensorBase> getTensors(const IndexExpr& expr);
 struct AccessTensorNode : public AccessNode {
   AccessTensorNode(TensorBase tensor, const std::vector<IndexVar>& indices)
       :  AccessNode(tensor.getTensorVar(), indices, {}, false), 
-         tensor(tensor) {}
+         tensorPtr(tensor.content) {}
 
   AccessTensorNode(TensorBase tensor, const std::vector<std::shared_ptr<IndexVarInterface>>& indices)
-    : AccessNode(tensor.getTensorVar()), tensor(tensor) {
+    : AccessNode(tensor.getTensorVar()), tensorPtr(tensor.content) {
     // Create the vector of IndexVar to assign to this->indexVars.
     std::vector<IndexVar> ivars(indices.size());
     for (size_t i = 0; i < indices.size(); i++) {
@@ -517,10 +517,21 @@ struct AccessTensorNode : public AccessNode {
     this->indexVars = std::move(ivars);
   }
 
-  TensorBase tensor;
+  // We hold a weak_ptr to the accessed TensorBase to avoid creating a reference
+  // cycle between the accessed TensorBase and this AccessTensorNode, since the
+  // TensorBase will store the AccessTensorNode (as part of an IndexExpr) as a
+  // field on itself. Not using a weak pointer results in leaking TensorBases.
+  std::weak_ptr<TensorBase::Content> tensorPtr;
+  TensorBase getTensor() const {
+    TensorBase tensor;
+    tensor.content = tensorPtr.lock();
+    return tensor;
+  }
+
   virtual void setAssignment(const Assignment& assignment) {
-    tensor.syncDependentTensors();
+    auto tensor = this->getTensor();
 
+    tensor.syncDependentTensors();
     Assignment assign = makeReductionNotation(assignment);
 
     tensor.setNeedsPack(false);
@@ -751,7 +762,7 @@ static inline map<TensorVar, TensorBase> getTensors(const IndexExpr& expr) {
       taco_iassert(isa<AccessTensorNode>(node)) << "Unknown subexpression";
 
       if (!util::contains(arguments, node->tensorVar)) {
-        arguments.insert({node->tensorVar, to<AccessTensorNode>(node)->tensor});
+        arguments.insert({node->tensorVar, to<AccessTensorNode>(node)->getTensor()});
       }
 
       // Also add any tensors backing index sets of tensor accesses.
@@ -763,7 +774,7 @@ static inline map<TensorVar, TensorBase> getTensors(const IndexExpr& expr) {
       }
 
       // TODO (rohany): This seems like dead code.
-      TensorBase tensor = to<AccessTensorNode>(node)->tensor;
+      TensorBase tensor = to<AccessTensorNode>(node)->getTensor();
       if (!util::contains(inserted, tensor)) {
         inserted.insert(tensor);
         operands.push_back(tensor);