diff --git a/python/lsst/daf/butler/remote_butler/server/README.md b/python/lsst/daf/butler/remote_butler/server/README.md
new file mode 100644
index 0000000000..3188ba9750
--- /dev/null
+++ b/python/lsst/daf/butler/remote_butler/server/README.md
@@ -0,0 +1,57 @@
+# Butler server
+
+## Concurrency and caching
+
+The server internally uses `DirectButler` instances to retrieve data from the
+database.  `DirectButler` instances are not thread-safe, so we need a separate
+instance for each request.  It is expensive to create a `DirectButler` instance
+from scratch, so we create a "template" `DirectButler` the first a repository
+is accessed, and clone it for each request.  (The cloning process is managed by
+`LabeledButlerFactory`.)
+
+Within `DirectButler`, there are a number of internal caches.  Some of these
+caches assume that no external processes will be mutating the repository during
+the lifetime of the `DirectButler` instance, and lock in state at the first
+time the data is accessed.  This behavior is OK in the context of a single HTTP
+request in Butler server. However, it is problematic across requests if the
+repository can be changing in the background, because new requests wouldn't be
+able to see updated data.
+
+It is expected that the main data release repositories will be immutable,
+which would allow for more aggressive caching, but we don't yet have a way to
+configure "mutable" vs "immutable" repositories in the server.
+
+### Summary of caches that exist in `DirectButler`
+
+Caches that are shared globally between all Butler instances within a process:
+
+- `DimensionUniverse`
+- `DimensionGroup`
+- `StorageClassFactory`
+
+Caches shared between instances cloned from the same parent Butler instance:
+
+- Most SqlAlchemy `Table` objects created during managers' `initialize()` classmethods
+- Tables created by `register()` in `ByNameOpaqueTableStorageManager`
+- Dataset "tag" and "calib" tables created by
+  `ByDimensionsDatasetRecordStorageManagerUUID` and stored in `DatasetTypeCache`
+
+Caches copied between instances at the time of cloning:
+
+- `DimensionRecordCache`
+- `DatasetType` and `DynamicTables` instances in `DatasetTypeCache`
+
+Caches that start empty in a newly cloned instance:
+
+- Collection cache and collection summary cache in `CachingContext`
+- `DimensionGraph` cache in `_DimensionGroupStorage`
+- `DatastoreCacheManager` in `FileDatastore` (not relevant to Butler server,
+  since the server does not access files on the filesystem.)
+
+### Caching caveats
+
+There is not currently a way to detect all changes to a Butler repository and invalidate the caches.  The Butler server must be restarted if a repository is changed in any of the following ways:
+
+- Any additions/deletions/modifications to governor dimension records
+- Updating a dataset type (deleting it and re-registering the same name with different values)
+- Deleting a dataset type