diff --git a/src/Lean/Meta/Basic.lean b/src/Lean/Meta/Basic.lean
index 2b4f5a4da288..3a437d4dcb7b 100644
--- a/src/Lean/Meta/Basic.lean
+++ b/src/Lean/Meta/Basic.lean
@@ -376,6 +376,19 @@ We should also investigate the impact on memory consumption.
 -/
 abbrev DefEqCache := PersistentHashMap DefEqCacheKey Bool
 
+/--
+A `DefEqTransCache` is a `DefEqCache` that is only valid in the `MetavarContext` in which it was declared.
+To keep track of whether this cache is still valid, it stores the `numAssignments` from the original `MetavarContext`.
+If the `numAssignments` in the `MetavarContext` has increased, we invalidate this cache.
+And when reverting the metavariable context in `checkpointDefEq`, if the `numAssignments`
+in the original `MetavarContext` is smaller than in this cache, we also invalidate this cache,
+and revert to the original cache.
+-/
+structure DefEqTransCache where
+  cache : DefEqCache := {}
+  numAssignments : Nat := 0
+  deriving Inhabited
+
 /--
 Cache datastructures for type inference, type class resolution, whnf, and definitional equality.
 -/
@@ -384,7 +397,7 @@ structure Cache where
   funInfo        : FunInfoCache := {}
   synthInstance  : SynthInstanceCache := {}
   whnf           : WhnfCache := {}
-  defEqTrans     : DefEqCache := {} -- transient cache for terms containing mvars or using nonstandard configuration options, it is frequently reset.
+  defEqTrans     : DefEqTransCache := {} -- transient cache for terms containing mvars or using nonstandard configuration options, it is valid as long as the count matches `MetavarContext.numAssignments`.
   defEqPerm      : DefEqCache := {} -- permanent cache for terms not containing mvars and using standard configuration options
   deriving Inhabited
 
@@ -574,10 +587,15 @@ instance : AddMessageContext MetaM where
 protected def saveState : MetaM SavedState :=
   return { core := (← Core.saveState), «meta» := (← get) }
 
-/-- Restore backtrackable parts of the state. -/
-def SavedState.restore (b : SavedState) : MetaM Unit := do
+/-- Restore backtrackable parts of the state.
+If `transCache == true`, then also reset the tranient defEq cache -/
+def SavedState.restore (b : SavedState) (transCache : Bool := false) : MetaM Unit := do
   b.core.restore
-  modify fun s => { s with mctx := b.meta.mctx, zetaDeltaFVarIds := b.meta.zetaDeltaFVarIds, postponed := b.meta.postponed }
+  modify fun s => { s with
+    mctx := b.meta.mctx
+    zetaDeltaFVarIds := b.meta.zetaDeltaFVarIds
+    postponed := b.meta.postponed
+    cache.defEqTrans := if transCache then b.meta.cache.defEqTrans else s.cache.defEqTrans }
 
 @[specialize, inherit_doc Core.withRestoreOrSaveFull]
 def withRestoreOrSaveFull (reusableResult? : Option (α × SavedState)) (act : MetaM α) :
@@ -639,8 +657,12 @@ def resetCache : MetaM Unit :=
 @[inline] def modifyInferTypeCache (f : InferTypeCache → InferTypeCache) : MetaM Unit :=
   modifyCache fun ⟨ic, c1, c2, c3, c4, c5⟩ => ⟨f ic, c1, c2, c3, c4, c5⟩
 
-@[inline] def modifyDefEqTransientCache (f : DefEqCache → DefEqCache) : MetaM Unit :=
-  modifyCache fun ⟨c1, c2, c3, c4, defeqTrans, c5⟩ => ⟨c1, c2, c3, c4, f defeqTrans, c5⟩
+/-- Modify the defEq transient cache. If it is not valid anymore, reset it before modifying it. -/
+@[inline] def modifyDefEqTransientCache (numAssignments : Nat) (f : DefEqCache → DefEqCache) : MetaM Unit :=
+  modifyCache fun c =>
+    let ⟨transCache, numAssignmentsOld⟩ := c.defEqTrans
+    let transCache := if numAssignments == numAssignmentsOld then transCache else {}
+    { c with defEqTrans := ⟨f transCache, numAssignments⟩ }
 
 @[inline] def modifyDefEqPermCache (f : DefEqCache → DefEqCache) : MetaM Unit :=
   modifyCache fun ⟨c1, c2, c3, c4, c5, defeqPerm⟩ => ⟨c1, c2, c3, c4, c5, f defeqPerm⟩
@@ -658,6 +680,9 @@ def mkDefEqCacheKey (lhs rhs : Expr) : MetaM DefEqCacheKey := do
 def mkInfoCacheKey (expr : Expr) (nargs? : Option Nat) : MetaM InfoCacheKey :=
   return { expr, nargs?, configKey := (← read).configKey }
 
+@[inline] def resetDefEqTransientCache : MetaM Unit :=
+  modify fun s => { s with cache.defEqTrans := ⟨{}, s.mctx.numAssignments⟩ }
+
 @[inline] def resetDefEqPermCaches : MetaM Unit :=
   modifyDefEqPermCache fun _ => {}
 
@@ -2349,16 +2374,6 @@ partial def processPostponed (mayPostpone : Bool := true) (exceptionOnFailure :=
 -/
 @[specialize] def checkpointDefEq (x : MetaM Bool) (mayPostpone : Bool := true) : MetaM Bool := do
   let s ← saveState
-  /-
-    It is not safe to use the `isDefEq` cache between different `isDefEq` calls.
-    Reason: different configuration settings, and result depends on the state of the `MetavarContext`
-    We have tried in the past to track when the result was independent of the `MetavarContext` state
-    but it was not effective. It is more important to cache aggressively inside of a single `isDefEq`
-    call because some of the heuristics create many similar subproblems.
-    See issue #1102 for an example that triggers an exponential blowup if we don't use this more
-    aggressive form of caching.
-  -/
-  modifyDefEqTransientCache fun _ => {}
   let postponed ← getResetPostponed
   try
     if (← x) then
@@ -2367,10 +2382,14 @@ partial def processPostponed (mayPostpone : Bool := true) (exceptionOnFailure :=
         setPostponed (postponed ++ newPostponed)
         return true
       else
-        s.restore
+        -- The transient cache needs to be reverted if it assumes an assignments that is being reverted.
+        let isInvalidCache := s.meta.mctx.numAssignments != (← get).cache.defEqTrans.numAssignments
+        s.restore (transCache := isInvalidCache)
         return false
     else
-      s.restore
+      -- The transient cache needs to be reverted if it assumes an assignments that is being reverted.
+      let isInvalidCache := s.meta.mctx.numAssignments != (← get).cache.defEqTrans.numAssignments
+      s.restore (transCache := isInvalidCache)
       return false
   catch ex =>
     s.restore
@@ -2405,6 +2424,16 @@ def isExprDefEq (t s : Expr) : MetaM Bool :=
     Remark: the kernel does *not* update the type of variables in the local context.
     -/
     resetDefEqPermCaches
+    /-
+      It is not safe to use the transient `isDefEq` cache between different `isDefEq` calls.
+      Reason: different configuration settings, and result depends on the state of the `MetavarContext`
+      We have tried in the past to track when the result was independent of the `MetavarContext` state
+      but it was not effective. It is more important to cache aggressively inside of a single `isDefEq`
+      call because some of the heuristics create many similar subproblems.
+      See issue #1102 and `tests/lean/run/defEqTransCache.lean` for examples that trigger an exponential blowup
+      if we don't use this more aggressive form of caching.
+    -/
+    resetDefEqTransientCache
     checkpointDefEq (mayPostpone := true) <| Meta.isExprDefEqAux t s
 
 /--
diff --git a/src/Lean/Meta/ExprDefEq.lean b/src/Lean/Meta/ExprDefEq.lean
index 1c3975c8949c..456d57212bfa 100644
--- a/src/Lean/Meta/ExprDefEq.lean
+++ b/src/Lean/Meta/ExprDefEq.lean
@@ -80,7 +80,7 @@ where
       if !isStructureLike (← getEnv) ctorVal.induct then
         trace[Meta.isDefEq.eta.struct] "failed, type is not a structure{indentExpr b}"
         return false
-      else if (← isDefEq (← inferType a) (← inferType b)) then
+      else if (← checkpointDefEq <| Meta.isExprDefEqAux (← inferType a) (← inferType b)) then
         checkpointDefEq do
           let args := b.getAppArgs
           let params := args[*...ctorVal.numParams].toArray
@@ -95,7 +95,7 @@ where
                 -- See comment at `isAbstractedUnassignedMVar`.
                 continue
             trace[Meta.isDefEq.eta.struct] "{a} =?= {b} @ [{j}], {proj} =?= {args[i]}"
-            unless (← isDefEq proj args[i]) do
+            unless (← Meta.isExprDefEqAux proj args[i]) do
               trace[Meta.isDefEq.eta.struct] "failed, unexpected arg #{i}, projection{indentExpr proj}\nis not defeq to{indentExpr args[i]}"
               return false
           return true
@@ -1240,7 +1240,7 @@ private partial def processAssignment (mvarApp : Expr) (v : Expr) : MetaM Bool :
   ```
 -/
 private def processAssignment' (mvarApp : Expr) (v : Expr) : MetaM Bool := do
-  if (← processAssignment mvarApp v) then
+  if (← checkpointDefEq <| processAssignment mvarApp v) then
     return true
   else
     let vNew ← whnf v
@@ -1248,7 +1248,7 @@ private def processAssignment' (mvarApp : Expr) (v : Expr) : MetaM Bool := do
       if mvarApp == vNew then
         return true
       else
-        processAssignment mvarApp vNew
+        checkpointDefEq <| processAssignment mvarApp vNew
     else
       return false
 
@@ -1629,7 +1629,7 @@ private def isDefEqProofIrrel (t s : Expr) : MetaM LBool := do
 private def isDefEqMVarSelf (mvar : Expr) (args₁ args₂ : Array Expr) : MetaM Bool := do
   if args₁.size != args₂.size then
     pure false
-  else if (← isDefEqArgs mvar args₁ args₂) then
+  else if (← checkpointDefEq <| isDefEqArgs mvar args₁ args₂) then
     pure true
   else if !(← isAssignable mvar) then
     pure false
@@ -1861,8 +1861,8 @@ end
 
 private def isDefEqOnFailure (t s : Expr) : MetaM Bool := do
   withTraceNodeBefore `Meta.isDefEq.onFailure (return m!"{t} =?= {s}") do
-    unstuckMVar t (fun t => Meta.isExprDefEqAux t s) <|
-    unstuckMVar s (fun s => Meta.isExprDefEqAux t s) <|
+    unstuckMVar t (fun t => checkpointDefEq <| Meta.isExprDefEqAux t s) <|
+    unstuckMVar s (fun s => checkpointDefEq <| Meta.isExprDefEqAux t s) <|
     tryUnificationHints t s <||> tryUnificationHints s t
 
 /--
@@ -1940,11 +1940,11 @@ private def isDefEqProj : Expr → Expr → MetaM Bool
   | .proj m i t, .proj n j s => do
     if (← read).inTypeClassResolution then
       -- See comment at `inTypeClassResolution`
-      pure (i == j && m == n) <&&> Meta.isExprDefEqAux t s
+      pure (i == j && m == n) <&&> checkpointDefEq (Meta.isExprDefEqAux t s)
     else if !backward.isDefEq.lazyProjDelta.get (← getOptions) then
-      pure (i == j && m == n) <&&> Meta.isExprDefEqAux t s
+      pure (i == j && m == n) <&&> checkpointDefEq (Meta.isExprDefEqAux t s)
     else if i == j && m == n then
-      isDefEqProjDelta t s i
+      checkpointDefEq (isDefEqProjDelta t s i)
     else
       return false
   | .proj structName 0 s, v  => isDefEqSingleton structName s v
@@ -2061,12 +2061,12 @@ private def isExprDefEqExpensive (t : Expr) (s : Expr) : MetaM Bool := do
       isDefEqOnFailure t s
 
 inductive DefEqCacheKind where
-  | transient -- problem has mvars or is using nonstandard configuration, we should use transient cache
+  | transient (numAssignments : Nat) -- problem has mvars or is using nonstandard configuration, we should use transient cache
   | permanent -- problem does not have mvars and we are using standard config, we can use one persistent cache.
 
 private def getDefEqCacheKind (t s : Expr) : MetaM DefEqCacheKind := do
   if t.hasMVar || s.hasMVar || (← read).canUnfold?.isSome then
-    return .transient
+    return .transient (← getMCtx).numAssignments
   else
     return .permanent
 
@@ -2084,7 +2084,12 @@ private def mkCacheKey (t s : Expr) : MetaM DefEqCacheKeyInfo := do
 
 private def getCachedResult (keyInfo : DefEqCacheKeyInfo) : MetaM LBool := do
   let cache ← match keyInfo.kind with
-    | .transient => pure (← get).cache.defEqTrans
+    | .transient numAssignments =>
+      let ⟨cache, numAssignmentsCache⟩ := (← get).cache.defEqTrans
+      if numAssignments == numAssignmentsCache then
+        pure cache
+      else
+        return .undef
     | .permanent => pure (← get).cache.defEqPerm
   match cache.find? keyInfo.key with
   | some val => return val.toLBool
@@ -2094,14 +2099,17 @@ private def cacheResult (keyInfo : DefEqCacheKeyInfo) (result : Bool) : MetaM Un
   let key := keyInfo.key
   match keyInfo.kind with
   | .permanent => modifyDefEqPermCache fun c => c.insert key result
-  | .transient =>
+  | .transient numAssignmentsOld =>
     /-
-    We must ensure that all assigned metavariables in the key are replaced by their current assignments.
-    Otherwise, the key is invalid after the assignment is "backtracked".
-    See issue #1870 for an example.
+    If the result is `false`, we cache it at `numAssignmentsOld`.
+    If the result is `true`, we only cache it if the number of assignments hasn't increased.
     -/
-    let key ← mkDefEqCacheKey (← instantiateMVars key.lhs) (← instantiateMVars key.rhs)
-    modifyDefEqTransientCache fun c => c.insert key result
+    if !result then
+      modifyDefEqTransientCache numAssignmentsOld fun c => c.insert key result
+    else
+      let numAssignmentsNew := (← getMCtx).numAssignments
+      if numAssignmentsOld == numAssignmentsNew then
+        modifyDefEqTransientCache numAssignmentsOld fun c => c.insert key result
 
 private def whnfCoreAtDefEq (e : Expr) : MetaM Expr := do
   if backward.isDefEq.lazyWhnfCore.get (← getOptions) then
diff --git a/src/Lean/MetavarContext.lean b/src/Lean/MetavarContext.lean
index d79cae6eda8d..af50f67b3a71 100644
--- a/src/Lean/MetavarContext.lean
+++ b/src/Lean/MetavarContext.lean
@@ -349,6 +349,8 @@ structure MetavarContext where
   lAssignment    : PersistentHashMap LMVarId Level := {}
   /-- Assignment table for expression metavariables.-/
   eAssignment    : PersistentHashMap MVarId Expr := {}
+  /-- Number of assignments in `eAssignment` and `lAssignments`. -/
+  numAssignments : Nat := 0
   /-- Assignment table for delayed abstraction metavariables.
   For more information about delayed abstraction, see the docstring for `DelayedMetavarAssignment`. -/
   dAssignment    : PersistentHashMap MVarId DelayedMetavarAssignment := {}
@@ -507,8 +509,9 @@ def hasAssignableMVar [Monad m] [MonadMCtx m] : Expr → m Bool
   This is a low-level API, and it is safer to use `isLevelDefEq (mkLevelMVar mvarId) u`.
 -/
 def assignLevelMVar [MonadMCtx m] (mvarId : LMVarId) (val : Level) : m Unit :=
-  modifyMCtx fun m => { m with lAssignment := m.lAssignment.insert mvarId val }
+  modifyMCtx fun m => { m with lAssignment := m.lAssignment.insert mvarId val, numAssignments := m.numAssignments + 1 }
 
+-- `assignLevelMVarExp` is only used in `instantiateMVars`, so it doesn't need to increment `numAssignments`
 @[export lean_assign_lmvar]
 def assignLevelMVarExp (m : MetavarContext) (mvarId : LMVarId) (val : Level) : MetavarContext :=
   { m with lAssignment := m.lAssignment.insert mvarId val }
@@ -520,8 +523,9 @@ a cycle is being introduced, or whether the expression has the right type.
 This is a low-level API, and it is safer to use `isDefEq (mkMVar mvarId) x`.
 -/
 def _root_.Lean.MVarId.assign [MonadMCtx m] (mvarId : MVarId) (val : Expr) : m Unit :=
-  modifyMCtx fun m => { m with eAssignment := m.eAssignment.insert mvarId val }
+  modifyMCtx fun m => { m with eAssignment := m.eAssignment.insert mvarId val, numAssignments := m.numAssignments + 1 }
 
+-- `assignExp` is only used in `instantiateMVars`, so it doesn't need to increment `numAssignments`
 @[export lean_assign_mvar]
 def assignExp (m : MetavarContext) (mvarId : MVarId) (val : Expr) : MetavarContext :=
   { m with eAssignment := m.eAssignment.insert mvarId val }
diff --git a/tests/lean/run/defEqTransCache.lean b/tests/lean/run/defEqTransCache.lean
new file mode 100644
index 000000000000..40d297c7d4cb
--- /dev/null
+++ b/tests/lean/run/defEqTransCache.lean
@@ -0,0 +1,77 @@
+import Lean
+/-!
+Previously, unification wouldn't be very careful with the `isDefEq` cache for terms containing metavariables.
+- This is mostly problematic because erasing the cache leads to exponential slowdowns (`test1` & `test2`)
+- but in some cases it lead to metavariable assignments leaking into places where they shouldn't be,
+  which either caused unification to fail where it should succeed (`test3`)
+  or to succeed where it is expected to fail (which happened in one mathlib proof).
+-/
+
+set_option maxHeartbeats 1000
+
+namespace test1
+class A (n : Nat) where
+  x : Nat
+
+instance [A n] : A (n+1) where
+  x := A.x n
+
+theorem test [A 0] : A.x 100 = 0 := sorry
+
+-- This rewrite should fail. Previously, it failed exponentially slowly
+example [A 1] : A.x 100 = 0 := by
+  fail_if_success rw [@test]
+  sorry
+end test1
+
+
+namespace test2
+@[irreducible] def A : Type := Unit
+
+@[irreducible] def B : Type := Unit
+
+unseal B in
+@[coe] def AtoB (_a : A) : B := ()
+
+instance : Coe A B where coe := AtoB
+
+def h {α : Type} (a b : α) : Nat → α
+| 0 => a
+| n + 1 => h b a n
+
+def f {α : Type} (a b : α) : Nat → Prop
+| 0 => a = b
+| n + 1 => f (h a b n) (h b a n) n ∧ f (h a a n) (h b b n) n
+
+axiom foo {p} {α : Type} (a b : α) : f a b p
+
+variable (x : A) (y : B)
+-- Previously, this check was exponentially slow; now it is quadratically slow
+#check (foo (↑x) y : f (AtoB x) y 30)
+end test2
+
+
+namespace test3
+structure A (α : Type) where
+  x : Type
+  y : α
+
+structure B (α : Type) extends A α where
+  z : Nat
+
+def A.map {α β} (f : α → β) (a : A α) : A β := ⟨a.x, f a.y⟩
+
+open Lean Meta in
+elab "unfold_head" e:term : term => do
+  let e ← Elab.Term.elabTerm e none
+  unfoldDefinition e
+
+-- use `unfold_head` to get the raw kernel projection `·.1` instead of the projection funtcion `A.x`
+def test {α} (i : B α) : unfold_head i.toA.x := sorry
+
+-- Previously, in this example the unification failed,
+-- because some metavariable assignment wasn't reverted properly
+-- However, it is clearly the case that `(i.toA.map f).x` is the same as `i.toA.x`
+example (i : B α) (f : α → β) : (i.toA.map f).x := by
+  apply @test
+end test3