diff --git a/navi/config.py b/navi/config.py
index 5b06595..38f48ce 100644
--- a/navi/config.py
+++ b/navi/config.py
@@ -59,7 +59,7 @@
     # Context compression
     context_compression_enabled: bool = True
     context_compression_threshold: float = 0.80   # trigger at 80% of ollama_num_ctx
-    context_keep_recent: int = 10                  # conversational turns to keep verbatim
+    context_keep_recent: int = 8                   # conversational turns to keep verbatim
     context_summary_temperature: float = 0.3
 
     # Global personality prompt prepended to every agent's system prompt.
diff --git a/navi/core/compressor.py b/navi/core/compressor.py
index 9560003..383ed7d 100644
--- a/navi/core/compressor.py
+++ b/navi/core/compressor.py
@@ -18,10 +18,12 @@
 
 _SUMMARIZE_SYSTEM = (
     "You are summarizing a conversation history to free up context space. "
-    "Produce a concise factual summary covering: key facts the user shared, "
-    "decisions made, tasks completed or in progress, important outputs or findings. "
-    "Use bullet points. Be brief — this summary replaces the conversation in context. "
-    "Do not include greetings or filler."
+    "Produce a structured factual summary covering: key facts the user shared, "
+    "decisions made, tasks completed or in progress, important outputs or findings, "
+    "any code or config snippets that were produced. "
+    "Use bullet points grouped by topic. Be thorough — capture enough detail that "
+    "the assistant can continue the conversation without the original messages. "
+    "Do not include greetings, filler, or meta-commentary about the summary itself."
 )
 
 
@@ -119,7 +121,7 @@
 
 # Safety limit: truncate formatted input to this many characters before sending to LLM.
 # Prevents the summarizer from receiving near-context-sized input it can't fit alongside output.
-_MAX_SUMMARY_INPUT_CHARS = 12_000
+_MAX_SUMMARY_INPUT_CHARS = 24_000
 
 
 async def compress_context(