diff --git a/litellm-proxy-extras/litellm_proxy_extras/migrations/20260423090000_memory_unique_nulls_not_distinct/migration.sql b/litellm-proxy-extras/litellm_proxy_extras/migrations/20260423090000_memory_unique_nulls_not_distinct/migration.sql
new file mode 100644
index 00000000000..3846d3961b6
--- /dev/null
+++ b/litellm-proxy-extras/litellm_proxy_extras/migrations/20260423090000_memory_unique_nulls_not_distinct/migration.sql
@@ -0,0 +1,34 @@
+-- Fix duplicate-key issue on LiteLLM_MemoryTable.
+--
+-- Problem: the unique constraint `(key, user_id, team_id)` does not block
+-- duplicates when `user_id` or `team_id` is NULL, because by default
+-- Postgres treats each NULL as distinct (ANSI SQL semantics). Callers with
+-- no team_id could POST the same key repeatedly and get multiple rows.
+--
+-- Fix (Postgres 15+): recreate the unique index with NULLS NOT DISTINCT so
+-- NULL values are treated as equal for uniqueness checks.
+--
+-- Note: Prisma has no schema syntax for NULLS NOT DISTINCT (as of writing),
+-- so this lives as a raw SQL migration. Prisma introspection will still see
+-- the index as uniquely covering `(key, user_id, team_id)`, matching the
+-- `@@unique` in schema.prisma.
+
+-- 1. Deduplicate existing rows: keep only the most-recently-updated row per
+--    (key, user_id, team_id), using IS NOT DISTINCT FROM so NULL == NULL.
+DELETE FROM "LiteLLM_MemoryTable" a
+USING "LiteLLM_MemoryTable" b
+WHERE a.memory_id <> b.memory_id
+  AND a.key = b.key
+  AND a.user_id IS NOT DISTINCT FROM b.user_id
+  AND a.team_id IS NOT DISTINCT FROM b.team_id
+  AND (
+    a.updated_at < b.updated_at
+    OR (a.updated_at = b.updated_at AND a.memory_id < b.memory_id)
+  );
+
+-- 2. Drop the old (NULL-distinct) unique index.
+DROP INDEX IF EXISTS "LiteLLM_MemoryTable_key_user_id_team_id_key";
+
+-- 3. Recreate with NULLS NOT DISTINCT.
+CREATE UNIQUE INDEX "LiteLLM_MemoryTable_key_user_id_team_id_key"
+    ON "LiteLLM_MemoryTable"("key", "user_id", "team_id") NULLS NOT DISTINCT;