diff --git a/kate/chat_renderer.cpp b/kate/chat_renderer.cpp
index fa552fc..b3de0b9 100644
--- a/kate/chat_renderer.cpp
+++ b/kate/chat_renderer.cpp
@@ -589,11 +589,25 @@ void ChatRenderer::render(QTextEdit *view, const QString &rawText)
     if (m_applyStorage)
         m_applyStorage->clear();
 
-    QList<ChatBlock> blocks = parseBlocks(rawText);
+    m_allBlocks = parseBlocks(rawText);
+
+    // Only render the tail window
+    m_renderStart = qMax(0, m_allBlocks.size() - m_windowSize);
+
     QString html = QStringLiteral(
         "<html><body style=\"font-family:sans-serif;font-size:medium;\">");
-    for (const auto &block : blocks)
-        html += blockToHtml(block);
+
+    // Indicate truncation if there are older blocks
+    if (m_renderStart > 0) {
+        html += QStringLiteral(
+            "<div id=\"ollie-load-more\" style=\"text-align:center;padding:8px;"
+            "color:gray;font-size:small;font-style:italic;\">↑ "
+            "Scroll up to load earlier messages (%1 more) ↑</div>")
+            .arg(m_renderStart);
+    }
+
+    for (int i = m_renderStart; i < m_allBlocks.size(); ++i)
+        html += blockToHtml(m_allBlocks[i]);
     html += QStringLiteral("</body></html>");
 
     auto *sb = view->verticalScrollBar();
@@ -619,7 +633,7 @@ void ChatRenderer::render(QTextEdit *view, const QString &rawText)
     sb->blockSignals(false);
 
     m_lastRawText = rawText;
-    m_lastBlockCount = blocks.size();
+    m_lastBlockCount = m_allBlocks.size();
 }
 
 bool ChatRenderer::renderIncremental(QTextEdit *view, const QString &rawText)
@@ -636,3 +650,49 @@ bool ChatRenderer::renderIncremental(QTextEdit *view, const QString &rawText)
     render(view, rawText);
     return true;
 }
+
+int ChatRenderer::prependOlder(QTextEdit *view, int count)
+{
+    if (m_renderStart <= 0)
+        return 0;
+
+    detectTheme();
+
+    int loadFrom = qMax(0, m_renderStart - count);
+    int loaded = m_renderStart - loadFrom;
+    if (loaded <= 0)
+        return 0;
+
+    m_renderStart = loadFrom;
+
+    // Re-render all currently-visible blocks (now including the older ones)
+    auto *sb = view->verticalScrollBar();
+    int oldMax = sb->maximum();
+    int oldVal = sb->value();
+
+    sb->blockSignals(true);
+
+    QString html = QStringLiteral(
+        "<html><body style=\"font-family:sans-serif;font-size:medium;\">");
+    if (m_renderStart > 0) {
+        html += QStringLiteral(
+            "<div id=\"ollie-load-more\" style=\"text-align:center;padding:8px;"
+            "color:gray;font-size:small;font-style:italic;\">↑ "
+            "Scroll up to load earlier messages (%1 more) ↑</div>")
+            .arg(m_renderStart);
+    }
+    for (int i = m_renderStart; i < m_allBlocks.size(); ++i)
+        html += blockToHtml(m_allBlocks[i]);
+    html += QStringLiteral("</body></html>");
+
+    view->setHtml(html);
+
+    // Restore scroll position offset by the added content height
+    int newMax = sb->maximum();
+    int addedHeight = newMax - oldMax;
+    sb->setValue(oldVal + addedHeight);
+
+    sb->blockSignals(false);
+
+    return loaded;
+}
diff --git a/kate/chat_renderer.h b/kate/chat_renderer.h
index 03bfc61..189b1a6 100644
--- a/kate/chat_renderer.h
+++ b/kate/chat_renderer.h
@@ -41,6 +41,10 @@ struct SubBlock {
 /**
  * ChatRenderer parses ollie chat output into structured blocks
  * and renders them as styled HTML in a QTextEdit.
+ *
+ * To avoid performance degradation on long conversations, only a
+ * tail window of blocks is eagerly rendered. Earlier blocks are
+ * loaded on demand when the user scrolls up.
  */
 class ChatRenderer
 {
@@ -62,13 +66,19 @@ public:
     };
     void setApplyStorage(QList<ApplyBlock> *storage) { m_applyStorage = storage; }
 
+    /// Maximum number of blocks to render eagerly (tail window).
+    static constexpr int DefaultWindowSize = 40;
+    void setWindowSize(int n) { m_windowSize = n; }
+    int windowSize() const { return m_windowSize; }
+
     /// Parse raw chat text into blocks
     QList<ChatBlock> parseBlocks(const QString &text) const;
 
     /// Parse a block's content into sub-blocks (prose, code, diff)
     QList<SubBlock> parseSubBlocks(const QString &content) const;
 
-    /// Render full chat text into the QTextEdit as styled HTML
+    /// Render the tail window of blocks into the QTextEdit.
+    /// Parses all blocks but only generates HTML for the last windowSize().
     void render(QTextEdit *view, const QString &rawText);
 
     /// Render incrementally: assumes rawText is the full text and only
@@ -76,6 +86,19 @@ public:
     /// an incremental update, false if it did a full re-render.
     bool renderIncremental(QTextEdit *view, const QString &rawText);
 
+    /// Prepend older blocks into the view. Returns the number of
+    /// blocks actually prepended (0 if all blocks are already shown).
+    int prependOlder(QTextEdit *view, int count);
+
+    /// True if there are older blocks not yet rendered.
+    bool hasOlderBlocks() const { return m_renderStart > 0; }
+
+    /// Total number of parsed blocks.
+    int totalBlocks() const { return m_allBlocks.size(); }
+
+    /// Number of blocks currently rendered.
+    int renderedBlocks() const { return m_allBlocks.size() - m_renderStart; }
+
 private:
     QString blockToHtml(const ChatBlock &block) const;
     QString subBlockToHtml(const SubBlock &sub, bool preserveWhitespace = false) const;
@@ -90,6 +113,11 @@ private:
     QString m_lastRawText; // for incremental rendering
     int m_lastBlockCount = 0;
 
+    // Windowed rendering state
+    QList<ChatBlock> m_allBlocks; // all parsed blocks
+    int m_renderStart = 0;        // index of first rendered block
+    int m_windowSize = DefaultWindowSize;
+
     // Colors (derived from palette)
     mutable bool m_dark = false;
     void detectTheme() const;
diff --git a/kate/ollie_kate.cpp b/kate/ollie_kate.cpp
index 7d92b64..f393312 100644
--- a/kate/ollie_kate.cpp
+++ b/kate/ollie_kate.cpp
@@ -387,7 +387,7 @@ OllieKateView::OllieKateView(OllieKatePlugin *plugin, KTextEditor::MainWindow *m
     m_chatView->installEventFilter(this); // handle resize in eventFilter
     repositionBtn();
 
-    // Also show/hide on manual scroll
+    // Also show/hide on manual scroll + lazy-load older messages
     connect(m_chatView->verticalScrollBar(), &QScrollBar::valueChanged, this, [this](int value) {
         auto *sb = m_chatView->verticalScrollBar();
         if (value >= sb->maximum() - 4) {
@@ -395,6 +395,13 @@ OllieKateView::OllieKateView(OllieKatePlugin *plugin, KTextEditor::MainWindow *m
         } else if (!m_chatView->toPlainText().isEmpty()) {
             m_scrollBottomBtn->show();
         }
+
+        // Lazy-load: when scrolled near the top, prepend older blocks
+        if (value <= 50 && m_renderer->hasOlderBlocks() && !m_loadingOlder) {
+            m_loadingOlder = true;
+            m_renderer->prependOlder(m_chatView, 20);
+            m_loadingOlder = false;
+        }
     });
 
     // Prompt input
diff --git a/kate/ollie_kate.h b/kate/ollie_kate.h
index 85f8726..8836a85 100644
--- a/kate/ollie_kate.h
+++ b/kate/ollie_kate.h
@@ -146,6 +146,7 @@ private:
     void annotateFileReferences(const QString &chatText);
 
     QList<ChatRenderer::ApplyBlock> m_applyBlocks;
+    bool m_loadingOlder = false; // guard against recursive scroll-triggered loads
 
     // Inline annotations (MovingRanges highlighting lines the agent references)
     QList<KTextEditor::MovingRange *> m_annotations;