diff --git a/navi/memory/extractor.py b/navi/memory/extractor.py
index 9fbfbc0..5069941 100644
--- a/navi/memory/extractor.py
+++ b/navi/memory/extractor.py
@@ -17,30 +17,41 @@
 log = structlog.get_logger()
 
 _EXTRACT_SYSTEM = """\
-You extract stable facts about the user from a conversation transcript.
+You extract stable facts about the user from a session transcript.
 
-Extract ONLY facts that are:
+The transcript contains four types of entries:
+1. User messages — what the user explicitly said
+2. Assistant messages — Navi's own responses
+3. Tool calls — which tools Navi executed and their arguments
+4. Tool results — the raw output those tools returned (may be truncated)
+
+Facts discovered through tool results are MORE RELIABLE than conversation.
+If a tool result contradicts something said in chat — trust the tool result.
+
+Extract facts that are:
 - Persistent characteristics: name, age, location, occupation, family situation
-- Technical environment: OS, tools, languages, home servers, devices
-- Preferences: communication style, coding habits, things they like or dislike
+- Technical environment: OS, tools, servers, devices, IPs, running services
+- Preferences: communication style, coding habits, likes/dislikes, workflow patterns
 - Ongoing projects or goals
-- Any other stable, reusable facts about this specific person
 
 Do NOT extract:
 - Topics that were discussed or questions that were asked
 - Temporary states ("was tired", "was busy today")
 - Information about third parties that isn't about the user
-- Directory-specific project notes, one-off commands, file paths, task progress, or local quirks
-  that belong in NAVI.md instead of long-term user memory
-- Duplicate facts already present in the transcript
+- Directory-specific project notes, one-off commands, file paths, task progress
+- NAVI.md content or local operational notes
+- Already-known facts that appear in the transcript
+
+For each fact, indicate its source:
+- "conversation" — the user explicitly stated it or it was inferred from chat
+- "tool_call" — discovered through a tool execution result
 
 Return ONLY a valid JSON array. No markdown, no prose, no comments.
 Return empty [] if nothing new should be extracted.
 Schema:
 [
-  {"category": "profile", "key": "name", "value": "Eugene"},
-  {"category": "technical", "key": "primary_os", "value": "Arch Linux"},
-  {"category": "preferences", "key": "response_language", "value": "Russian"}
+  {"category": "profile", "key": "name", "value": "Eugene", "source": "conversation", "source_context": "user introduced themselves"},
+  {"category": "technical", "key": "host_ip", "value": "192.168.1.168", "source": "tool_call", "source_context": "found via terminal ip addr"}
 ]
 
 Valid categories: profile, preferences, technical, projects, other"""
@@ -74,20 +85,49 @@
         await _regenerate_summary(llm, model, memory_store)
 
 
+_MAX_TOOL_RESULT_LEN = 500
+_MAX_TRANSCRIPT_CHARS = 12_000
+
+
 async def _extract_facts(session, llm: LLMBackend, model: str, store: MemoryStore) -> int:
     lines: list[str] = []
+
+    # Map tool_call_id -> tool_name so we can label tool results
+    tool_call_map: dict[str, str] = {}
+    for msg in session.messages:
+        if msg.role == "assistant" and msg.tool_calls:
+            for tc in msg.tool_calls:
+                tool_call_map[tc.id] = tc.name
+
     for msg in session.messages:
         if msg.role == "user" and msg.content:
             lines.append(f"User: {msg.content}")
-        elif msg.role == "assistant" and msg.content:
-            lines.append(f"Assistant: {msg.content}")
+        elif msg.role == "assistant":
+            if msg.content:
+                lines.append(f"Assistant: {msg.content}")
+            if msg.tool_calls:
+                for tc in msg.tool_calls:
+                    args = str(tc.arguments)[:200]
+                    lines.append(f"[Tool call] {tc.name}({args})")
+        elif msg.role == "tool" and msg.content:
+            tool_name = tool_call_map.get(msg.tool_call_id or "", "unknown")
+            content = msg.content
+            if len(content) > _MAX_TOOL_RESULT_LEN:
+                content = content[:_MAX_TOOL_RESULT_LEN] + " ... [truncated]"
+            lines.append(f"[Tool result] {tool_name}: {content}")
 
     if not lines:
         return 0
 
+    transcript = "\n".join(lines)
+    if len(transcript) > _MAX_TRANSCRIPT_CHARS:
+        # Keep early context + recent tail; drop the middle
+        half = _MAX_TRANSCRIPT_CHARS // 2
+        transcript = transcript[:half] + "\n\n... [transcript truncated] ...\n\n" + transcript[-half:]
+
     prompt = [
         Message(role="system", content=_EXTRACT_SYSTEM),
-        Message(role="user", content="\n".join(lines)),
+        Message(role="user", content=transcript),
     ]
 
     try:
@@ -118,14 +158,27 @@
             category = "other"
         key = str(fact.get("key", "")).strip()
         value = str(fact.get("value", "")).strip()
+        source = str(fact.get("source", "conversation")).strip().lower()
+        if source not in {"conversation", "tool_call", "auto_discovery", "user_explicit"}:
+            source = "conversation"
+        source_context = str(fact.get("source_context", "")).strip()
+
+        # Confidence mapping based on source reliability
+        confidence = {"tool_call": 95, "auto_discovery": 95, "user_explicit": 90}.get(source, 70)
+
         if key and value:
+            # TODO: Semantic deduplication — before upsert, search for semantically
+            # similar facts and merge/update instead of creating duplicates.
+            # Problem: LLM generates different keys for the same fact across sessions.
+            # Solution: vector search + similarity threshold before upsert.
             await store.upsert_fact(
                 category=category,
                 key=key,
                 value=value,
                 source_session_id=session.id,
-                source="conversation",
-                confidence=70,
+                source=source,
+                confidence=confidence,
+                source_context=source_context,
             )
             count += 1