diff --git a/openviking/prompts/templates/compression/memory_extraction.yaml b/openviking/prompts/templates/compression/memory_extraction.yaml
index 704726b89..ee097145c 100644
--- a/openviking/prompts/templates/compression/memory_extraction.yaml
+++ b/openviking/prompts/templates/compression/memory_extraction.yaml
@@ -204,9 +204,10 @@ template: |
 
   # Three-Level Structure
 
-  Each memory contains three levels, each serving a purpose:
+  Each memory contains three levels. **Keep all levels concise — memories are sticky notes, not essays.**
 
-  **abstract (L0)**: Index layer, plain text one-liner
+  **abstract (L0)**: Index layer, plain text one-liner. **Target: 1 sentence, ~50-80 characters.**
+  - This is the PRIMARY retrieval key — it MUST be specific enough to distinguish this memory from others.
   - Merge types (preferences/entities/profile/patterns): `[Merge key]: [Description]`
     - preferences: `Python code style: No type hints, concise and direct`
     - entities: `OpenViking project: AI Agent long-term memory management system`
@@ -216,13 +217,16 @@ template: |
     - events: `Decided to refactor memory system: Simplify to 5 categories`
     - cases: `Band not recognized → Request member/album/style details`
 
-  **overview (L1)**: Structured summary layer, organized with Markdown headings
+  **overview (L1)**: Structured summary layer, organized with Markdown headings. **Target: 3-5 bullet points.**
   - preferences: `## Preference Domain` / `## Specific Preferences`
   - entities: `## Basic Info` / `## Core Attributes`
   - events: `## Decision Content` / `## Reason` / `## Result`
   - cases: `## Problem` / `## Solution`
 
-  **content (L2)**: Detailed expansion layer, free Markdown, includes background, timeline, complete narrative
+  **content (L2)**: Core facts layer. **Target: 2-4 sentences with all essential specifics.**
+  - Capture ONLY the facts that would be lost if this memory were deleted.
+  - Include: names, versions, numbers, configurations, error messages, solutions.
+  - Exclude: background narratives, general explanations, elaboration of obvious points.
 
   # Few-shot Examples
 
diff --git a/openviking/prompts/templates/compression/memory_merge_bundle.yaml b/openviking/prompts/templates/compression/memory_merge_bundle.yaml
index 14bc77698..d1ab862b1 100644
--- a/openviking/prompts/templates/compression/memory_merge_bundle.yaml
+++ b/openviking/prompts/templates/compression/memory_merge_bundle.yaml
@@ -1,8 +1,8 @@
 metadata:
   id: "compression.memory_merge_bundle"
   name: "Memory Merge Bundle"
-  description: "Merge memory and return L0/L1/L2 in one structured response"
-  version: "1.0.0"
+  description: "Merge memory and return L0/L1/L2 in one structured response, with facet guard and length limits"
+  version: "2.0.0"
   language: "en"
   category: "compression"
 
@@ -46,7 +46,7 @@ variables:
     default: "auto"
 
 template: |
-  You are merging one existing memory with one new memory update.
+  You are deciding whether to merge two memories of the same category, or skip (keep them separate).
 
   Category: {{ category }}
   Target Output Language: {{ output_language }}
@@ -61,26 +61,62 @@ template: |
   - Overview (L1): {{ new_overview }}
   - Content (L2): {{ new_content }}
 
-  Requirements:
-  - Merge into a single coherent memory.
-  - Keep non-conflicting details from existing memory.
-  - Update conflicting details to reflect the newer fact.
-  - Output language must be {{ output_language }}.
-  - Return JSON only.
+  ## Step 1: Facet coherence check
+
+  Before merging, determine whether the two memories describe the SAME specific facet/topic.
+
+  Same facet examples (should merge):
+  - "Python code style: no type hints" + "Python code style: concise comments" → same facet (Python code style)
+  - "OpenViking project: memory extraction" + "OpenViking project: added dedup" → same facet (OpenViking project)
+
+  Different facet examples (should skip):
+  - "Python code style: no type hints" + "Food preference: likes apples" → different facets
+  - "Server 192.168.2.75: runs agent-helper" + "OpenViking project: memory extraction" → different facets
+  - "Git commit style: zh-CN verbs" + "Exit code semantics: 0/1/2" → different facets
+
+  If the two memories cover DIFFERENT facets, you MUST output decision "skip".
+  Do NOT merge unrelated information just because they share the same category.
+
+  ## Step 2: Merge (only if same facet)
+
+  If merging:
+  - When facts conflict, keep the NEWER version only.
+  - Condense to essential facts. Do NOT accumulate every historical detail.
+  - The merged memory should read as a clean, up-to-date snapshot — not a changelog.
+
+  ## Hard length limits
 
-  Output JSON schema:
+  - `abstract`: ≤ 80 characters
+  - `overview`: ≤ 200 characters
+  - `content`: ≤ 300 characters
+
+  If the merged result would exceed these limits, aggressively summarize.
+  Drop older, less important details first. Preserve specific values (names, numbers, versions) over narrative.
+
+  ## Output
+
+  Return JSON only. Two possible decisions:
+
+  When merging:
   {
     "decision": "merge",
-    "abstract": "one-line L0 summary",
-    "overview": "structured markdown L1 summary",
-    "content": "full merged L2 content",
+    "abstract": "one-line L0 (≤80 chars)",
+    "overview": "structured L1 (≤200 chars)",
+    "content": "condensed L2 (≤300 chars)",
     "reason": "short reason"
   }
 
+  When skipping (different facets):
+  {
+    "decision": "skip",
+    "reason": "short reason why these are different facets"
+  }
+
   Constraints:
   - `abstract` must be concise and specific.
-  - `overview` and `content` must be non-empty.
+  - `overview` and `content` must be non-empty when decision is "merge".
   - Do not output any text outside JSON.
+  - Output language must be {{ output_language }}.
 
 llm_config:
   temperature: 0.0
diff --git a/openviking/session/memory_extractor.py b/openviking/session/memory_extractor.py
index fefef0ba7..cff41e88f 100644
--- a/openviking/session/memory_extractor.py
+++ b/openviking/session/memory_extractor.py
@@ -474,7 +474,8 @@ async def create_memory(
                 owner_space=owner_space,
             )
             logger.info(f"uri {memory_uri} abstract: {payload.abstract} content: {payload.content}")
-            memory.set_vectorize(Vectorize(text=payload.content))
+            # Use abstract for vectorization — shorter text produces more discriminative embeddings
+            memory.set_vectorize(Vectorize(text=payload.abstract or payload.content))
             return memory
 
         # Determine parent URI based on category
@@ -514,7 +515,8 @@ async def create_memory(
             owner_space=owner_space,
         )
         logger.info(f"uri {memory_uri} abstract: {candidate.abstract} content: {candidate.content}")
-        memory.set_vectorize(Vectorize(text=candidate.content))
+        # Use abstract for vectorization — shorter text produces more discriminative embeddings
+        memory.set_vectorize(Vectorize(text=candidate.abstract or candidate.content))
         return memory
 
     async def _append_to_profile(