diff --git a/hybridse/using-zetasql.md b/hybridse/using-zetasql.md
index cd11a22..3f7ad23 100644
--- a/hybridse/using-zetasql.md
+++ b/hybridse/using-zetasql.md
@@ -38,7 +38,7 @@ ZetaSQL is a SQL Analyzer Framework from Google. We are considering using it as
 
 We do some survey on ZetaSQL. ZetaSQL provides many APIs and services. 
 
-![zetasql analyzer workflow](./images/image-zetasql-workflow.png)
+![zetasql analyzer workflow](../images/image-zetasql-workflow.png)
 
 The figure above demonstrate the zetasql's analyze workflow. But we **only** use its parser module, since It can fullfill our requirements.
 
@@ -163,7 +163,7 @@ ZetaSQL supports standard ANSI SQL.  We are going to integrate zetasql's parser
 2. Transform `ASTStatement` into `LogicalPlan` .
 3. Convert `ASTExpression`  to `ExprNode` 
 
-![Figure2-use-zetasql-parser](./images/image-hybridse-new-parser.png)
+![Figure2-use-zetasql-parser](../images/image-hybridse-new-parser.png)
 
 **Pros**:
 
diff --git a/hybridse/images/image-hybridse-new-parser.png b/images/image-hybridse-new-parser.png
similarity index 100%
rename from hybridse/images/image-hybridse-new-parser.png
rename to images/image-hybridse-new-parser.png
diff --git a/hybridse/images/image-zetasql-workflow.png b/images/image-zetasql-workflow.png
similarity index 100%
rename from hybridse/images/image-zetasql-workflow.png
rename to images/image-zetasql-workflow.png
diff --git a/images/memory_pool.png b/images/memory_pool.png
new file mode 100644
index 0000000..7403676
Binary files /dev/null and b/images/memory_pool.png differ
diff --git a/fedb/api-server.md b/openmldb/api-server.md
similarity index 100%
rename from fedb/api-server.md
rename to openmldb/api-server.md
diff --git a/openmldb/memory-pool.md b/openmldb/memory-pool.md
new file mode 100644
index 0000000..7aa4644
--- /dev/null
+++ b/openmldb/memory-pool.md
@@ -0,0 +1,84 @@
+- Start Date: 2021-06-24
+- Target Major Version: (2.4.0)
+- Reference Issues: 
+- Implementation PR: 
+
+# Memory Pool
+
+# Summary
+OpenMLDB use [TCMalloc](https://github.com/gperftools/gperftools) as memory alloction lib currently. Like other mallocs, the heap managed by TCMalloc consists of a set of pages. A run of contiguous pages is represented by a Span object. A span can either be allocated, or free. But if only a small buffer in a span is not freed, this span can not release to os which is called memory fragment. The memory is already located in central free list when a partition in OpenMLDB is dropped.
+
+So We try to design a novel memory pool to avoid this problem.
+
+# Detailed Design
+
+When insert one record to OpenMLDB, it will call multiple alloctions and construct structs/classes as follows: 
+| data structure | size | description
+| --- | --- | --- |
+| key | key length | 
+| value | value length | 
+| DataBlock | 16 |
+| KeyEntry | 40 | 
+| Node<Slice, void*>| 40 |
+| Node<uint64_t, void*>| 32 | 
+| std::atomic<Node<Slice, std::string*>*> node0[12] | 96 | 8 * height
+
+By analyzing the above data structure, we found that the memory alloction has the following situations and we can use different alloction strategy.
+
+| - | - | struct / class | alloction strategy |
+| --- | --- | --- | --- |
+| Normal | Fixed length | KeyEntry | boost::pool
+| Normal | Variable length| Key | Tcmalloc/Var-length pool
+| Timeserise | Fixed length | Node | TimeSerisePool
+| Timeserise | Variable length | Value | TimeSerisePool
+
+
+TimeSerisePool is composed of multiple TimeBuckets as followers. 
+
+<div align=center><img src="../images/memory_pool.png"/></div>
+
+```C++
+class TimeBucket {
+  public:
+    struct Block {
+        Block* next;
+        char data[]; 	    
+    };
+    Clear(); // free the block list
+    void* Alloc(uint32_t size) {
+        object_num_++;    
+        if (current_offset_ + size <= block_size_ - sizeof(Block)) {
+            void* addr = head_->data + current_offset_;
+            current_offset_ += size;
+            return addr;
+        } else {
+            Block* block = (Block*)malloc(block_size_);
+            current_offset_ = size;
+            block->next = head_->next;
+            head_ = block;
+            return head_->data;
+        }
+    }
+  private:
+    uint32_t block_size_;
+    uint32_t current_offset_;
+    uint32_t object_num_;
+    Block* head_;  
+};
+```
+
+```C++
+class TimeSerisePool {
+  public:
+    Alloc(uint32_t size, uint64_t time);
+    Free(void*, uint64_t time);
+  private:
+    // key is the time / (60 * 60 * 1000)
+    std::map<uint32_t, std::unique_ptr<TimeBucket>> pool_;
+};
+```
+
+Note: TimeSerisePool does not guarantee thread safety.
+
+# Adoption stratege
+If we implement this proposal, there is no impact on existing interfaces
\ No newline at end of file