diff --git a/src/daemon/embedded_daemon.cpp b/src/daemon/embedded_daemon.cpp
index 68c97cf..0e83df9 100644
--- a/src/daemon/embedded_daemon.cpp
+++ b/src/daemon/embedded_daemon.cpp
@@ -591,16 +591,23 @@ bool EmbeddedDaemon::start(const std::string& binary_path)
         args.push_back("-maxconnections=" + std::to_string(max_connections_));
     }
 
+    // The user's own DRAGONX.conf wins over our flags (a command-line flag would override it): read the
+    // conf once, for a normal start only (an isolated start has its own datadir + conf and keeps the pin).
+    std::string conf;
+    if (override_datadir_.empty()) {
+        std::ifstream in((fs::path(util::Platform::getDragonXDataDir()) / "DRAGONX.conf").string(), std::ios::binary);
+        if (in) conf.assign(std::istreambuf_iterator<char>(in), std::istreambuf_iterator<char>());
+        if (confSetsKey(conf, "dbcache")) {
+            DEBUG_LOGF("[INFO] DRAGONX.conf sets dbcache; not pinning -dbcache\n");
+        }
+        dropPinnedDbcacheIfConfSets(args, conf);
+    }
+
     // More RPC workers for a node we launch: dragonxd's default 8 workers + 16-deep queue fill up with
     // requests waiting behind a long cs_main holder (a witness rebuild), after which even "stop" is refused.
     // The GUI itself no longer floods it; this is headroom for dragonx-cli and other clients — a delay,
     // not a fix. Not for an isolated start, and DRAGONX.conf wins (a flag would override the user's own).
     if (override_datadir_.empty()) {
-        std::string conf;
-        {
-            std::ifstream in((fs::path(util::Platform::getDragonXDataDir()) / "DRAGONX.conf").string(), std::ios::binary);
-            if (in) conf.assign(std::istreambuf_iterator<char>(in), std::istreambuf_iterator<char>());
-        }
         if (!confSetsKey(conf, "rpcthreads")) args.push_back("-rpcthreads=16");
     }
 
diff --git a/src/daemon/embedded_daemon.h b/src/daemon/embedded_daemon.h
index d341b72..47be8b1 100644
--- a/src/daemon/embedded_daemon.h
+++ b/src/daemon/embedded_daemon.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <algorithm>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
@@ -424,6 +425,18 @@ public:
         return false;
     }
 
+    /**
+     * Drop the wallet's own -dbcache= pin from `args` when the user's DRAGONX.conf sets dbcache (a
+     * command-line flag overrides the conf, so the user's own value must win). Pure; no logging.
+     */
+    static void dropPinnedDbcacheIfConfSets(std::vector<std::string>& args, const std::string& conf)
+    {
+        if (!confSetsKey(conf, "dbcache")) return;
+        args.erase(std::remove_if(args.begin(), args.end(),
+                       [](const std::string& a) { return a.rfind("-dbcache=", 0) == 0; }),
+                   args.end());
+    }
+
     /** Decision returned by evaluateDatadirLockGate(): whether start() may spawn now. */
     struct StartLockGateDecision {
         bool proceed = true;             // false => bail before spawning
diff --git a/tests/test_phase4.cpp b/tests/test_phase4.cpp
index 34597ef..3a10869 100644
--- a/tests/test_phase4.cpp
+++ b/tests/test_phase4.cpp
@@ -6859,6 +6859,25 @@ void testConfSetsKey()
     EXPECT_FALSE(EmbeddedDaemon::confSetsKey("rpcworkqueue=64 # rpcthreads=4\n", "rpcthreads"));
 }
 
+// C97: a user dbcache= in DRAGONX.conf wins over the wallet's -dbcache pin; default argv is unchanged.
+void testDropPinnedDbcache()
+{
+    using dragonx::daemon::EmbeddedDaemon;
+    const std::vector<std::string> base{"-a", "-dbcache=450", "-b"};
+    auto run = [&](const std::string& conf) {
+        std::vector<std::string> v = base;
+        EmbeddedDaemon::dropPinnedDbcacheIfConfSets(v, conf);
+        return v;
+    };
+    const std::vector<std::string> dropped{"-a", "-b"};
+    EXPECT_TRUE(run("dbcache=1000") == dropped);
+    EXPECT_TRUE(run("  dbcache = 8\r\n") == dropped);
+    EXPECT_TRUE(run("rpcuser=x\ndbcache=2000\n") == dropped);
+    EXPECT_TRUE(run("#dbcache=1") == base);
+    EXPECT_TRUE(run("") == base);
+    EXPECT_TRUE(run("rpcuser=x") == base);
+}
+
 // S12: the ownership record names the node we launched, by pid AND start time, so a later session can
 // recognise it after "Leave running" — and a reused pid, or a record from before a reboot, never matches.
 void testOwnershipRecord()
@@ -18655,6 +18674,7 @@ int main()
     testDatadirLockGate();
     testDaemonProcessIdentity();
     testConfSetsKey();
+    testDropPinnedDbcache();
     testOwnershipRecord();
     testDebugLogTailRules();
     testNoteConsolidation();