diff --git a/crates/tinymold-remote/src/client/mod.rs b/crates/tinymold-remote/src/client/mod.rs
index c3adc7a..3d898ad 100644
--- a/crates/tinymold-remote/src/client/mod.rs
+++ b/crates/tinymold-remote/src/client/mod.rs
@@ -60,7 +60,12 @@ impl Client {
     pub(crate) fn new(config: &RemoteConfig) -> Result<Self> {
         let mut endpoint = Endpoint::from_shared(config.endpoint.clone())
             .map_err(Error::Connect)?
-            .connect_timeout(config.timeout);
+            .connect_timeout(config.timeout)
+            // With hyper's fixed windows (2 MiB a stream, 5 MiB a
+            // connection), a read can move at most one window per round
+            // trip, however fast the link; adaptive windows grow to the
+            // bandwidth-delay product.
+            .http2_adaptive_window(true);
         if config.uses_tls() {
             endpoint = endpoint
                 .tls_config(ClientTlsConfig::new().with_webpki_roots())
diff --git a/docs/specs/remote.md b/docs/specs/remote.md
index d115c2d..3f95f1b 100644
--- a/docs/specs/remote.md
+++ b/docs/specs/remote.md
@@ -94,7 +94,9 @@ sharing a key.
    outcome is `remote`.
 
 Calls go through ByteStream `Read` for blobs, so blob size is not limited
-by the gRPC message size. Once an entry's signature checks out, its
+by the gRPC message size. The connection uses HTTP/2 adaptive flow-control
+windows, so reads from a distant cache are not held to a fixed window per
+round trip. Once an entry's signature checks out, its
 objects and its record are read together, within the same limits as a
 push's uploads (below), and a pre-key
 lookup reads its signature and record together: each read is a round trip,