diff --git a/openai_server/agent_prompting.py b/openai_server/agent_prompting.py
index 0fd72ac47..43ddc07d7 100644
--- a/openai_server/agent_prompting.py
+++ b/openai_server/agent_prompting.py
@@ -588,6 +588,79 @@ def get_audio_transcription_helper():
         audio_transcription = ''
     return audio_transcription
 
+def get_save_memory_helper():
+    cwd = os.path.abspath(os.getcwd())
+    save_solution_memor = f"""\n
+* You always keep an eye on the errors you encounter and the solutions you find to them.
+* You have to save the solutions to memory so that you can refer to them in the future and avoid making the same mistakes again.
+* You can also save new memories to make sure you don't forget them and make use of them in the future.
+* Never skips saving the error solutions to the memory.
+
+* Save error solutions to memory using python. Use for saving error solutions to memory.
+* For saving an error solution to memory, you are recommended to use the existing pre-built python code, E.g.:
+```sh
+# filename: my_save_solution_memory.sh
+# execution: true
+python {cwd}/openai_server/agent_tools/save_memory.py --task "TASK" --error "ERROR" --solution "SOLUTION"
+
+* usage: python {cwd}/openai_server/agent_tools/save_memory.py [-h] --task "TASK" --error "ERROR" --solution "SOLUTION"
+* You should save solutions you have found to errors while solving user tasks. 
+* Solutions have to be callable codes if possible, otherwise just put explanations.
+* While saving the solution, you should explicityl mention: 1-the task that lead you to the error,
+2-the error you encountered, and 3-the solution you found to the error, as a code or explanation.
+* Example task: 'While trying to scrape X data from the web I used the 123.xyz URL but it was blocked by the server.'
+* Example error: 'Error 403: Forbidden'
+* Example solution: 'For similar type of data, I found this another URL 456.xyz that worked.'
+* Another example solution: 'Use following code to scrape X data from the web: ...'
+* It's really important to save the solutions to memory so that you can refer to them in the future and avoid making the same mistakes again.
+"""
+    return save_solution_memor
+
+def get_memories(instruction:str):
+    # read all the csv files that starts with the name 'memory_' in the directory: openai_files/62224bfb-c832-4452-81e7-8a4bdabbe164/
+    # and concatenate them into single memory_df
+
+    # find memory paths via os
+    memory_df_paths = []
+    # TODO: This is just a toy code. In real usage, the memory files should be stored in a stable DB
+    for root, dirs, files in os.walk('openai_files/62224bfb-c832-4452-81e7-8a4bdabbe164/'):
+        for file in files:
+            if file.startswith('memory_') and file.endswith('.csv'):
+                memory_df_paths.append(os.path.join(root, file))
+    print(f"Memory Paths: {memory_df_paths}")
+    # if no memory files found, return empty string
+    if len(memory_df_paths) == 0:
+        return ""
+
+    from openai_server.agent_utils import MemoryVectorDB
+    # Initialize vector DB with OpenAI model
+    # TODO: In the real usage, there has to be a stable vectordb tha will work accross different chats
+    # Currently this is just a dummy vectordb to test the functionality
+    memory_db = MemoryVectorDB(
+        model="text-embedding-3-small",
+        openai_api_key=ast.literal_eval(os.getenv('H2OGPT_H2OGPT_API_KEYS'))[0],
+        openai_base_url="https://api.gpt.h2o.ai/v1"
+        )
+
+    import pandas as pd
+    memory_df = pd.concat([pd.read_csv(memory_df_path) for memory_df_path in memory_df_paths])
+    # Create VectorDB documents from memory_df rows
+    documents = []
+    for index, row in memory_df.iterrows():
+        document = f"{row['task']}, {row['error']}, {row['solution']}"
+        documents.append(document)
+    # Add documents to VectorDB
+    memory_db.add_texts(documents)
+
+    # Get the most similar 5 documents to the instruction
+    results, distances = memory_db.query(instruction, k=5, threshold=0.95)
+    if len(results) == 0:
+        return ""
+
+    memory_prompt = "\n# Previous Solutions Memory:"
+    # join results with new line, also add index to each result
+    memory_prompt += "\n".join([f"Memory-{i+1}:\n {result}" for i, result in enumerate(results)])
+    return memory_prompt
 
 def get_full_system_prompt(agent_code_writer_system_message, agent_system_site_packages, system_prompt, base_url,
                            api_key, model, text_context_list, image_file, temp_dir, query):
@@ -598,6 +671,9 @@ def get_full_system_prompt(agent_code_writer_system_message, agent_system_site_p
     mermaid_renderer_helper = get_mermaid_renderer_helper()
     image_generation_helper = get_image_generation_helper()
     audio_transcription_helper = get_audio_transcription_helper()
+    save_memory_helper = get_save_memory_helper()
+    memories_prompt = get_memories(query)
+    print(f"Memories Prompt: {memories_prompt}")
 
     chat_doc_query, internal_file_names = get_chat_doc_context(text_context_list, image_file,
                                                                temp_dir,
@@ -614,6 +690,6 @@ def get_full_system_prompt(agent_code_writer_system_message, agent_system_site_p
 
     agent_tools_note = f"\nDo not hallucinate agent_tools tools. The only files in the {path_agent_tools} directory are as follows: {list_dir}\n"
 
-    system_message = agent_code_writer_system_message + image_query_helper + mermaid_renderer_helper + image_generation_helper + audio_transcription_helper + agent_tools_note + chat_doc_query
+    system_message = agent_code_writer_system_message + image_query_helper + mermaid_renderer_helper + image_generation_helper + audio_transcription_helper + save_memory_helper + memories_prompt + agent_tools_note + chat_doc_query
     # TODO: Also return image_generation_helper and audio_transcription_helper ? 
     return system_message, internal_file_names, chat_doc_query, image_query_helper, mermaid_renderer_helper
diff --git a/openai_server/agent_tools/save_memory.py b/openai_server/agent_tools/save_memory.py
new file mode 100644
index 000000000..132974229
--- /dev/null
+++ b/openai_server/agent_tools/save_memory.py
@@ -0,0 +1,21 @@
+import argparse
+import pandas as pd
+import uuid
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Save new memory to be used for future references")
+    parser.add_argument("--task", type=str, required=True, help="Detailed task explanation that lead to the error or the solution")
+    parser.add_argument("--error", type=str, required=True, help="Error message that was encountered, if there was any. ")
+    parser.add_argument("--solution", type=str, required=True, help="Solution, always includes codes. Full method codes are preferred so that the solution can be recalled as is.")
+    args = parser.parse_args()
+    # Memory file
+    memory_file = f"memory_{str(uuid.uuid4())[:6]}.csv"
+    # new memory
+    memory = pd.DataFrame([[args.task, args.error, args.solution]], columns=['task', 'error', 'solution'])
+    # write the memory back to the file
+    memory.to_csv(memory_file, index=False)
+    print(f"New memory saved: {memory_file}")
+
+if __name__ == "__main__":
+    main()
diff --git a/openai_server/agent_utils.py b/openai_server/agent_utils.py
index 50ddabd8c..e61fdf68a 100644
--- a/openai_server/agent_utils.py
+++ b/openai_server/agent_utils.py
@@ -304,3 +304,100 @@ def get_ret_dict_and_handle_files(chat_result, temp_dir, agent_verbose, internal
     ret_dict.update(dict(temp_dir=temp_dir))
 
     return ret_dict
+
+
+import faiss
+import numpy as np
+from openai import OpenAI
+class MemoryVectorDB:
+    def __init__(self, model: str, openai_base_url:str, openai_api_key: str):
+        # Initialize OpenAI embeddings model
+        self.client = OpenAI(base_url=openai_base_url, api_key=openai_api_key)
+        self.model = model
+
+        # Initialize FAISS index (using L2 distance)
+        self.index = None
+        self.texts = []
+        self.embeddings = None
+        self.id_map = {}
+
+    def get_embeddings(self, texts: list):
+        # Generate embedding  for the texts via client
+        response = self.client.embeddings.create(
+        input=texts,
+        model=self.model
+    )
+        # To reach embeddings for the first item in the list, use response.data[0].embedding and so on
+        embeddings = []
+        for i in range(len(response.data)):
+            embeddings.append(response.data[i].embedding)
+
+        embedding_matrix = np.array(embeddings).astype('float32')
+        return embedding_matrix
+
+    def add_texts(self, texts: list):
+        # Generate embeddings for the texts
+        embedding_matrix = self.get_embeddings(texts)
+
+        # Update the list of stored texts and id map
+        start_id = len(self.texts)
+        self.texts.extend(texts)
+        for i, text in enumerate(texts):
+            self.id_map[start_id + i] = text
+
+        # Create or update the FAISS index
+        if self.index is None:
+            # Initialize the FAISS index with the embedding dimension
+            self.index = faiss.IndexFlatL2(embedding_matrix.shape[1])
+            # Add embeddings to the FAISS index
+            self.index.add(embedding_matrix)
+            self.embeddings = embedding_matrix
+        else:
+            self.index.add(embedding_matrix)
+            self.embeddings = np.vstack((self.embeddings, embedding_matrix))
+
+        # Confirm embeddings were added
+        print("Texts added successfully.")
+        print("Number of items in FAISS index:", self.index.ntotal)
+
+    def query(self, query_text: str, k: int = 2, threshold: float = 2.0):
+        # Generate embedding for the query
+        query_embedding_np = self.get_embeddings([query_text])
+
+        # Check if FAISS index is initialized
+        if self.index is None or self.index.ntotal == 0:
+            raise ValueError("FAISS index is empty or not initialized. Please add texts before querying.")
+
+        # Perform FAISS search
+        D, I = self.index.search(query_embedding_np, k)
+        
+        # Ensure valid indices and handle potential errors
+        results = []
+        distances = []
+        for i, idx in enumerate(I[0]):
+            if idx in self.id_map:
+                results.append(self.id_map[idx])
+                distances.append(D[0][i])
+            else:
+                print(f"Warning: Index {idx} not found in id_map. It might have been deleted.")
+        final_results = [r for r, d in zip(results, distances) if d <= threshold]
+        final_distances = [d for d in distances if d <= threshold]
+        print(f"Memory VetorDB: Returns {len(final_results)} results.")
+        # Returns results having distance less than or equal to the threshold
+        return final_results, final_distances
+
+    def delete_text_by_id(self, idx: int):
+        # Remove the text from stored texts and id map
+        if idx in self.id_map:
+            del self.id_map[idx]
+            self.texts.pop(idx)
+
+            # Remove the embedding from FAISS index and rebuild the index
+            self.embeddings = np.delete(self.embeddings, idx, axis=0)
+            self.index.reset()
+            self.index.add(self.embeddings)
+        else:
+            print(f"Warning: Text with index {idx} not found in the database.")
+
+    def get_all_texts(self):
+        return self.texts