diff --git a/CLAUDE.md b/CLAUDE.md
index 6f8253f8b19..dcfda34f4e6 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -7,6 +7,7 @@ This file contains core rules for AI coding agents. For detailed patterns and wo
 **Core Documents (Read directly):**
 - `tests/CLAUDE.md` - Testing guidelines and strategies
 - `BUILDING.md` - Build system, quality enforcement, BUILD policies
+- `src/ai/backend/manager/models/alembic/README.md` - Alembic migration backport strategy
 - `README.md` - Project overview and architecture
 - `proposals/README.md` - BEP (Backend.AI Enhancement Proposals)
 
@@ -49,6 +50,12 @@ pants test --changed-since=HEAD~1
 
 **Fix all lint, type, and test errors — never suppress or skip.**
 
+## Alembic Migration Backport
+
+When backporting migrations to release branches, both the backport and main branch
+migrations must be idempotent. See `src/ai/backend/manager/models/alembic/README.md`
+for the full strategy and examples.
+
 ## Layer Architecture
 
 API Handler → Processor → Service → Repository → DB
diff --git a/changes/10607.doc.md b/changes/10607.doc.md
new file mode 100644
index 00000000000..d729b1db78f
--- /dev/null
+++ b/changes/10607.doc.md
@@ -0,0 +1 @@
+Add `ALEMBIC.md` documenting the Alembic migration backport strategy with idempotent writing rules and step-by-step procedure.
diff --git a/src/ai/backend/account_manager/models/alembic/README b/src/ai/backend/account_manager/models/alembic/README
index e0d0858f266..523ce67b8ea 100644
--- a/src/ai/backend/account_manager/models/alembic/README
+++ b/src/ai/backend/account_manager/models/alembic/README
@@ -1 +1,3 @@
-Generic single-database configuration with an async dbapi.
\ No newline at end of file
+Generic single-database configuration with an async dbapi.
+
+For the migration backport strategy, see `src/ai/backend/manager/models/alembic/README.md`.
diff --git a/src/ai/backend/appproxy/coordinator/models/alembic/README b/src/ai/backend/appproxy/coordinator/models/alembic/README
index 98e4f9c44ef..a33ac3de1fb 100644
--- a/src/ai/backend/appproxy/coordinator/models/alembic/README
+++ b/src/ai/backend/appproxy/coordinator/models/alembic/README
@@ -1 +1,3 @@
-Generic single-database configuration.
\ No newline at end of file
+Generic single-database configuration.
+
+For the migration backport strategy, see `src/ai/backend/manager/models/alembic/README.md`.
diff --git a/src/ai/backend/manager/models/alembic/CLAUDE.md b/src/ai/backend/manager/models/alembic/CLAUDE.md
new file mode 100644
index 00000000000..427ea88f3c5
--- /dev/null
+++ b/src/ai/backend/manager/models/alembic/CLAUDE.md
@@ -0,0 +1,16 @@
+# Alembic Migrations — Guardrails
+
+> For the full backport procedure and examples, see `README.md` in this directory.
+
+## Rules
+
+- **Backport = fixes only** -- Never backport feature-level schema changes.
+- **Idempotent** -- All backport migrations (both `d` and `d'`) must use existence
+  checks (`IF NOT EXISTS`, `IF EXISTS`, `inspector`) so they are safe to re-apply.
+- **Release version comment** -- Every migration file must have a
+  `# Part of: <major>.<minor>.<patch>` comment next to the revision identifiers.
+  For backports: `# Part of: 26.2.1 (backport), 26.3.0 (main)`.
+- **No revision editing on released migrations** -- Never modify `revision` or
+  `down_revision` of a migration that has already been released. Editing
+  `down_revision` is allowed only when inserting a backport into the main chain
+  before merge.
diff --git a/src/ai/backend/manager/models/alembic/README b/src/ai/backend/manager/models/alembic/README
deleted file mode 100644
index 98e4f9c44ef..00000000000
--- a/src/ai/backend/manager/models/alembic/README
+++ /dev/null
@@ -1 +0,0 @@
-Generic single-database configuration.
\ No newline at end of file
diff --git a/src/ai/backend/manager/models/alembic/README.md b/src/ai/backend/manager/models/alembic/README.md
new file mode 100644
index 00000000000..6f55fe0286a
--- /dev/null
+++ b/src/ai/backend/manager/models/alembic/README.md
@@ -0,0 +1,161 @@
+# Alembic Migrations
+
+Generic single-database configuration.
+
+## Migration Backport Strategy
+
+This section describes how to safely backport Alembic migrations to release branches
+while keeping the main branch consistent.
+
+The same strategy applies to all components:
+
+| Component | Alembic directory |
+|---|---|
+| Manager | `src/ai/backend/manager/models/alembic/` |
+| Account Manager | `src/ai/backend/account_manager/models/alembic/` |
+| App Proxy Coordinator | `src/ai/backend/appproxy/coordinator/models/alembic/` |
+
+### Principles
+
+1. **Fixes only** -- Backport migrations must contain schema fixes only (e.g., missing
+   columns, incorrect types, enum coexistence issues). Feature-level schema changes are
+   never backported.
+2. **Idempotent** -- Both the backport migration and its duplicate on main must be
+   idempotent so they are safe to re-apply on databases that may already have the change.
+3. **No manual revision editing** -- Never modify the `revision` or `down_revision` of
+   an existing migration that has already been released. Editing `down_revision` is
+   allowed only when inserting a backport migration into the main branch chain before
+   the change is merged (see Step 2 below).
+
+### Backport Procedure
+
+Given the following migration chain on **main**:
+
+```
+a  -->  b  -->  c        (a = backport target head, c = main head)
+```
+
+#### Step 1: Create migration `d` on the release branch
+
+On the release branch, create a new migration whose `down_revision` is `a` (the
+release branch head at the point you are targeting):
+
+```
+a  -->  d                 (release branch result)
+```
+
+#### Step 2: Insert `d` into main and add duplicate `d'`
+
+On **main**, the same migration `d` is inserted between `a` and `b`, and a duplicate
+`d'` is appended on top of the current main head `c`:
+
+```
+a  -->  d  -->  b  -->  c  -->  d'
+```
+
+- `d` — Set `down_revision = a` and `revision` to the same value used on the release
+  branch. Update `b`'s `down_revision` to point to `d`.
+- `d'` — A new migration file with a fresh revision ID, `down_revision = c`, that
+  performs the **same schema change** as `d` but written idempotently so it is a no-op
+  on databases that already applied `d`.
+
+#### Step 3: Merge the head if necessary
+
+After inserting `d` into the chain, run `alembic heads` on main. If there are multiple
+heads, create a merge migration:
+
+```bash
+alembic merge heads -m "merge backport head"
+```
+
+### Release Version Comment
+
+Every migration file must include a comment indicating which release version
+(including the minor version, e.g., `26.3.0`) it belongs to. Place the comment
+next to the revision identifiers:
+
+```python
+# revision identifiers, used by Alembic.
+revision = "1cc9b47e0a8e"
+down_revision = "ffcf0ed13a26"
+# Part of: 26.3.0
+branch_labels = None
+depends_on = None
+```
+
+For backport migrations, note both the target release branch and the main branch:
+
+```python
+# Part of: 26.2.1 (backport), 26.3.0 (main)
+```
+
+### Idempotent Writing Rules
+
+Every backport migration (both `d` and `d'`) **must** be idempotent. Use the following
+patterns:
+
+#### DDL guards
+
+```python
+# Creating a table
+conn = op.get_bind()
+inspector = sa.inspect(conn)
+if "my_table" not in inspector.get_table_names():
+    op.create_table("my_table", ...)
+
+# Adding a column
+columns = [c["name"] for c in inspector.get_columns("my_table")]
+if "new_col" not in columns:
+    op.add_column("my_table", sa.Column("new_col", sa.String))
+
+# Creating an index
+indexes = [idx["name"] for idx in inspector.get_indexes("my_table")]
+if "ix_my_table_col" not in indexes:
+    op.create_index("ix_my_table_col", "my_table", ["col"])
+```
+
+#### Enum type guards
+
+```python
+conn = op.get_bind()
+result = conn.exec_driver_sql(
+    "SELECT 1 FROM pg_type WHERE typname = 'myenum'"
+)
+if result.fetchone() is None:
+    # Create the enum type
+    my_enum = sa.Enum("A", "B", name="myenum")
+    my_enum.create(conn)
+```
+
+#### Raw SQL guards
+
+```sql
+-- Column
+ALTER TABLE my_table ADD COLUMN IF NOT EXISTS new_col TEXT;
+
+-- Index
+CREATE INDEX IF NOT EXISTS ix_my_table_col ON my_table (col);
+
+-- Dropping
+DROP INDEX IF EXISTS ix_my_table_col;
+ALTER TABLE my_table DROP COLUMN IF EXISTS old_col;
+```
+
+#### Downgrade
+
+Downgrade functions follow the same idempotent rules. If the downgrade is handled by
+another migration in the chain, use `pass`:
+
+```python
+def downgrade() -> None:
+    pass
+```
+
+### Real-World Examples
+
+See the following migrations in the codebase for reference:
+
+- `src/ai/backend/manager/models/alembic/versions/1cc9b47e0a8e_fix_sessionresult_enum_type_coexistence_backport.py`
+  -- Checks multiple possible enum states and fixes whichever scenario it finds.
+- `src/ai/backend/manager/models/alembic/versions/c4ea15b77136_ensure_auditlogs_table_exist.py`
+  -- Uses `inspector.get_table_names()` to skip table creation when it already exists.