{"id":731,"date":"2026-04-16T12:58:44","date_gmt":"2026-04-16T04:58:44","guid":{"rendered":"https:\/\/connectword.dpdns.org\/?p=731"},"modified":"2026-04-16T12:58:44","modified_gmt":"2026-04-16T04:58:44","slug":"how-to-build-a-universal-long-term-memory-layer-for-ai-agents-using-mem0-and-openai","status":"publish","type":"post","link":"https:\/\/connectword.dpdns.org\/?p=731","title":{"rendered":"How to Build a Universal Long-Term Memory Layer for AI Agents Using Mem0 and OpenAI"},"content":{"rendered":"

In this tutorial, we build a universal long-term memory layer for AI agents using Mem0<\/strong><\/a>, OpenAI models, and ChromaDB. We design a system that can extract structured memories from natural conversations, store them semantically, retrieve them intelligently, and integrate them directly into personalized agent responses. We move beyond simple chat history and implement persistent, user-scoped memory with full CRUD control, semantic search, multi-user isolation, and custom configuration. Finally, we construct a production-ready memory-augmented agent architecture that demonstrates how modern AI systems can reason with contextual continuity rather than operate statelessly.<\/p>\n

\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
!pip install mem0ai openai rich chromadb -q\n\n\nimport os\nimport getpass\nfrom datetime import datetime\n\n\nprint(\"=\" * 60)\nprint(\"\"\ud83d\udd10\"  MEM0 Advanced Tutorial \u2014 API Key Setup\")\nprint(\"=\" * 60)\n\n\nOPENAI_API_KEY = getpass.getpass(\"Enter your OpenAI API key: \")\nos.environ[\"OPENAI_API_KEY\"] = OPENAI_API_KEY\n\n\nprint(\"n\"\u2705\" API key set!n\")\n\n\nfrom openai import OpenAI\nfrom mem0 import Memory\nfrom rich.console import Console\nfrom rich.panel import Panel\nfrom rich.table import Table\nfrom rich.markdown import Markdown\nfrom rich import print as rprint\nimport json\n\n\nconsole = Console()\nopenai_client = OpenAI()\n\n\nconsole.rule(\"[bold cyan]MODULE 1: Basic Memory Setup[\/bold cyan]\")\n\n\nmemory = Memory()\n\n\nprint(Panel(\n   \"[green]\u2713 Memory instance created with default config[\/green]n\"\n   \"  \u2022 LLM: gpt-4.1-nano (OpenAI)n\"\n   \"  \u2022 Vector Store: ChromaDB (local)n\"\n   \"  \u2022 Embedder: text-embedding-3-small\",\n   title=\"Memory Config\", border_style=\"cyan\"\n))\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We install all required dependencies and securely configure our OpenAI API key. We initialize the Mem0 Memory instance along with the OpenAI client and Rich console utilities. We establish the foundation of our long-term memory system with the default configuration powered by ChromaDB and OpenAI embeddings.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
console.rule(\"[bold cyan]MODULE 2: Adding & Retrieving Memories[\/bold cyan]\")\n\n\nUSER_ID = \"alice_tutorial\"\n\n\nprint(\"n\"\ud83d\udcdd\" Adding memories for user:\", USER_ID)\n\n\nconversations = [\n   [\n       {\"role\": \"user\", \"content\": \"Hi! I'm Alice. I'm a software engineer who loves Python and machine learning.\"},\n       {\"role\": \"assistant\", \"content\": \"Nice to meet you Alice! Python and ML are great areas to be in.\"}\n   ],\n   [\n       {\"role\": \"user\", \"content\": \"I prefer dark mode in all my IDEs and I use VS Code as my main editor.\"},\n       {\"role\": \"assistant\", \"content\": \"Good to know! VS Code with dark mode is a popular combo.\"}\n   ],\n   [\n       {\"role\": \"user\", \"content\": \"I'm currently building a RAG pipeline for my company's internal docs. It's for a fintech startup.\"},\n       {\"role\": \"assistant\", \"content\": \"That's exciting! RAG pipelines are really valuable for enterprise use cases.\"}\n   ],\n   [\n       {\"role\": \"user\", \"content\": \"I have a dog named Max and I enjoy hiking on weekends.\"},\n       {\"role\": \"assistant\", \"content\": \"Max sounds lovely! Hiking is a great way to recharge.\"}\n   ],\n]\n\n\nresults = []\nfor i, convo in enumerate(conversations):\n   result = memory.add(convo, user_id=USER_ID)\n   extracted = result.get(\"results\", [])\n   for mem in extracted:\n       results.append(mem)\n   print(f\"  Conversation {i+1}: {len(extracted)} memory(ies) extracted\")\n\n\nprint(f\"n\"\u2705\" Total memories stored: {len(results)}\")<\/code><\/pre>\n<\/div>\n<\/div>\n
We simulate realistic multi-turn conversations and store them using Mem0\u2019s automatic memory extraction pipeline. We add structured conversational data for a specific user and allow the LLM to extract meaningful long-term facts. We verify how many memories are created, confirming that semantic knowledge is successfully persisted.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
console.rule(\"[bold cyan]MODULE 3: Semantic Search[\/bold cyan]\")\n\n\nqueries = [\n   \"What programming languages does the user prefer?\",\n   \"What is Alice working on professionally?\",\n   \"What are Alice's hobbies?\",\n   \"What tools and IDE does Alice use?\",\n]\n\n\nfor query in queries:\n   search_results = memory.search(query=query, user_id=USER_ID, limit=2)\n   table = Table(title=f\"\"\ud83d\udd0d\" Query: {query}\", show_lines=True)\n   table.add_column(\"Memory\", style=\"white\", max_width=60)\n   table.add_column(\"Score\", style=\"green\", justify=\"center\")\n  \n   for r in search_results.get(\"results\", []):\n       score = r.get(\"score\", \"N\/A\")\n       score_str = f\"{score:.4f}\" if isinstance(score, float) else str(score)\n       table.add_row(r[\"memory\"], score_str)\n  \n   console.print(table)\n   print()\n\n\nconsole.rule(\"[bold cyan]MODULE 4: CRUD Operations[\/bold cyan]\")\n\n\nall_memories = memory.get_all(user_id=USER_ID)\nmemories_list = all_memories.get(\"results\", [])\n\n\nprint(f\"n\"\ud83d\udcda\" All memories for '{USER_ID}':\")\nfor i, mem in enumerate(memories_list):\n   print(f\"  [{i+1}] ID: {mem['id'][:8]}...  \u2192  {mem['memory']}\")\n\n\nif memories_list:\n   first_id = memories_list[0][\"id\"]\n   original_text = memories_list[0][\"memory\"]\n  \n   print(f\"n\"\u270f\"  Updating memory: '{original_text}'\")\n   memory.update(memory_id=first_id, data=original_text + \" (confirmed)\")\n  \n   updated = memory.get(memory_id=first_id)\n   print(f\"   After update: '{updated['memory']}'\")<\/code><\/pre>\n<\/div>\n<\/div>\n
We perform semantic search queries to retrieve relevant memories using natural language. We demonstrate how Mem0 ranks stored memories by similarity score and returns the most contextually aligned information. We also perform CRUD operations by listing, updating, and validating stored memory entries.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
console.rule(\"[bold cyan]MODULE 5: Memory-Augmented Chat[\/bold cyan]\")\n\n\ndef chat_with_memory(user_message: str, user_id: str, session_history: list) -> str:\n  \n   relevant = memory.search(query=user_message, user_id=user_id, limit=5)\n   memory_context = \"n\".join(\n       f\"- {r['memory']}\" for r in relevant.get(\"results\", [])\n   ) or \"No relevant memories found.\"\n  \n   system_prompt = f\"\"\"You are a highly personalized AI assistant.\nYou have access to long-term memories about this user.\n\n\nRELEVANT USER MEMORIES:\n{memory_context}\n\n\nUse these memories to provide context-aware, personalized responses.\nBe natural \u2014 don't explicitly announce that you're using memories.\"\"\"\n  \n   messages = [{\"role\": \"system\", \"content\": system_prompt}]\n   messages.extend(session_history[-6:])\n   messages.append({\"role\": \"user\", \"content\": user_message})\n  \n   response = openai_client.chat.completions.create(\n       model=\"gpt-4.1-nano-2025-04-14\",\n       messages=messages\n   )\n   assistant_response = response.choices[0].message.content\n  \n   exchange = [\n       {\"role\": \"user\", \"content\": user_message},\n       {\"role\": \"assistant\", \"content\": assistant_response}\n   ]\n   memory.add(exchange, user_id=user_id)\n  \n   session_history.append({\"role\": \"user\", \"content\": user_message})\n   session_history.append({\"role\": \"assistant\", \"content\": assistant_response})\n  \n   return assistant_response\n\n\n\n\nsession = []\ndemo_messages = [\n   \"Can you recommend a good IDE setup for me?\",\n   \"What kind of project am I currently building at work?\",\n   \"Suggest a weekend activity I might enjoy.\",\n   \"What's a good tech stack for my current project?\",\n]\n\n\nprint(\"n\"\ud83e\udd16\" Starting memory-augmented conversation with Alice...n\")\n\n\nfor msg in demo_messages:\n   print(Panel(f\"[bold yellow]User:[\/bold yellow] {msg}\", border_style=\"yellow\"))\n   response = chat_with_memory(msg, USER_ID, session)\n   print(Panel(f\"[bold green]Assistant:[\/bold green] {response}\", border_style=\"green\"))\n   print()<\/code><\/pre>\n<\/div>\n<\/div>\n
We build a fully memory-augmented chat loop that retrieves relevant memories before generating responses. We dynamically inject personalized context into the system prompt and store each new exchange back into long-term memory. We simulate a multi-turn session to demonstrate contextual continuity and personalization in action.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
console.rule(\"[bold cyan]MODULE 6: Multi-User Memory Isolation[\/bold cyan]\")\n\n\nUSER_BOB = \"bob_tutorial\"\n\n\nbob_conversations = [\n   [\n       {\"role\": \"user\", \"content\": \"I'm Bob, a data scientist specializing in computer vision and PyTorch.\"},\n       {\"role\": \"assistant\", \"content\": \"Great to meet you Bob!\"}\n   ],\n   [\n       {\"role\": \"user\", \"content\": \"I prefer Jupyter notebooks over VS Code, and I use Vim keybindings.\"},\n       {\"role\": \"assistant\", \"content\": \"Classic setup for data science work!\"}\n   ],\n]\n\n\nfor convo in bob_conversations:\n   memory.add(convo, user_id=USER_BOB)\n\n\nprint(\"n\"\ud83d\udd10\" Testing memory isolation between Alice and Bob:n\")\n\n\ntest_query = \"What programming tools does this user prefer?\"\n\n\nalice_results = memory.search(query=test_query, user_id=USER_ID, limit=3)\nbob_results = memory.search(query=test_query, user_id=USER_BOB, limit=3)\n\n\nprint(\"\"\ud83d\udc69\" Alice's memories:\")\nfor r in alice_results.get(\"results\", []):\n   print(f\"   \u2022 {r['memory']}\")\n\n\nprint(\"n\"\ud83d\udc68\" Bob's memories:\")\nfor r in bob_results.get(\"results\", []):\n   print(f\"   \u2022 {r['memory']}\")<\/code><\/pre>\n<\/div>\n<\/div>\n
We demonstrate user-level memory isolation by introducing a second user with distinct preferences. We store separate conversational data and validate that searches remain scoped to the correct user_id. We confirm that memory namespaces are isolated, ensuring secure multi-user agent deployments.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
print(\"n\"\u2705\" Memory isolation confirmed \u2014 users cannot see each other's data.\")\n\n\nconsole.rule(\"[bold cyan]MODULE 7: Custom Configuration[\/bold cyan]\")\n\n\ncustom_config = {\n   \"llm\": {\n       \"provider\": \"openai\",\n       \"config\": {\n           \"model\": \"gpt-4.1-nano-2025-04-14\",\n           \"temperature\": 0.1,\n           \"max_tokens\": 2000,\n       }\n   },\n   \"embedder\": {\n       \"provider\": \"openai\",\n       \"config\": {\n           \"model\": \"text-embedding-3-small\",\n       }\n   },\n   \"vector_store\": {\n       \"provider\": \"chroma\",\n       \"config\": {\n           \"collection_name\": \"advanced_tutorial_v2\",\n           \"path\": \"\/tmp\/chroma_advanced\",\n       }\n   },\n   \"version\": \"v1.1\"\n}\n\n\ncustom_memory = Memory.from_config(custom_config)\n\n\nprint(Panel(\n   \"[green]\u2713 Custom memory instance created[\/green]n\"\n   \"  \u2022 LLM: gpt-4.1-nano with temperature=0.1n\"\n   \"  \u2022 Embedder: text-embedding-3-smalln\"\n   \"  \u2022 Vector Store: ChromaDB at \/tmp\/chroma_advancedn\"\n   \"  \u2022 Collection: advanced_tutorial_v2\",\n   title=\"Custom Config Applied\", border_style=\"magenta\"\n))\n\n\ncustom_memory.add(\n   [{\"role\": \"user\", \"content\": \"I'm a researcher studying neural plasticity and brain-computer interfaces.\"}],\n   user_id=\"researcher_01\"\n)\n\n\nresult = custom_memory.search(\"What field does this person work in?\", user_id=\"researcher_01\", limit=2)\nprint(\"n\"\ud83d\udd0d\" Custom memory search result:\")\nfor r in result.get(\"results\", []):\n   print(f\"   \u2022 {r['memory']}\")\n\n\nconsole.rule(\"[bold cyan]MODULE 8: Memory History[\/bold cyan]\")\n\n\nall_alice = memory.get_all(user_id=USER_ID)\nalice_memories = all_alice.get(\"results\", [])\n\n\ntable = Table(title=f\"\"\ud83d\udccb\" Full Memory Profile: {USER_ID}\", show_lines=True, width=90)\ntable.add_column(\"#\", style=\"dim\", width=3)\ntable.add_column(\"Memory ID\", style=\"cyan\", width=12)\ntable.add_column(\"Memory Content\", style=\"white\")\ntable.add_column(\"Created At\", style=\"yellow\", width=12)\n\n\nfor i, mem in enumerate(alice_memories):\n   mem_id = mem[\"id\"][:8] + \"...\"\n   created = mem.get(\"created_at\", \"N\/A\")\n   if created and created != \"N\/A\":\n       try:\n           created = datetime.fromisoformat(created.replace(\"Z\", \"+00:00\")).strftime(\"%m\/%d %H:%M\")\n       except:\n           created = str(created)[:10]\n   table.add_row(str(i+1), mem_id, mem[\"memory\"], created)\n\n\nconsole.print(table)\n\n\nconsole.rule(\"[bold cyan]MODULE 9: Memory Deletion[\/bold cyan]\")\n\n\nall_mems = memory.get_all(user_id=USER_ID).get(\"results\", [])\nif all_mems:\n   last_mem = all_mems[-1]\n   print(f\"n\"\ud83d\uddd1\"  Deleting memory: '{last_mem['memory']}'\")\n   memory.delete(memory_id=last_mem[\"id\"])\n  \n   updated_count = len(memory.get_all(user_id=USER_ID).get(\"results\", []))\n   print(f\"\"\u2705\" Deleted. Remaining memories for {USER_ID}: {updated_count}\")\n\n\nconsole.rule(\"[bold cyan]\"\u2705\" TUTORIAL COMPLETE[\/bold cyan]\")\n\n\nsummary = \"\"\"\n# \"\ud83c\udf93\" Mem0 Advanced Tutorial Summary\n\n\n## What You Learned:\n1. **Basic Setup** \u2014 Instantiate Memory with default & custom configs\n2. **Add Memories** \u2014 From conversations (auto-extracted by LLM)\n3. **Semantic Search** \u2014 Retrieve relevant memories by natural language query\n4. **CRUD Operations** \u2014 Get, Update, Delete individual memories\n5. **Memory-Augmented Chat** \u2014 Full pipeline: retrieve \u2192 respond \u2192 store\n6. **Multi-User Isolation** \u2014 Separate memory namespaces per user_id\n7. **Custom Configuration** \u2014 Custom LLM, embedder, and vector store\n8. **Memory History** \u2014 View full memory profiles with timestamps\n9. **Cleanup** \u2014 Delete specific or all memories\n\n\n## Key Concepts:\n- `memory.add(messages, user_id=...)`\n- `memory.search(query, user_id=...)`\n- `memory.get_all(user_id=...)`\n- `memory.update(memory_id, data)`\n- `memory.delete(memory_id)`\n- `Memory.from_config(config)`\n\n\n## Next Steps:\n- Swap ChromaDB for Qdrant, Pinecone, or Weaviate\n- Use the hosted Mem0 Platform (app.mem0.ai) for production\n- Integrate with LangChain, CrewAI, or LangGraph agents\n- Add `agent_id` for agent-level memory scoping\n\"\"\"\n\n\nconsole.print(Markdown(summary))<\/code><\/pre>\n<\/div>\n<\/div>\n
We create a fully custom Mem0 configuration with explicit parameters for the LLM, embedder, and vector store. We test the custom memory instance and explore memory history, timestamps, and structured profiling. Finally, we demonstrate deletion and cleanup operations, completing the full lifecycle management of long-term agent memory.<\/p>\n
In conclusion, we implemented a complete memory infrastructure for AI agents using Mem0 as a universal memory abstraction layer. We demonstrated how to add, retrieve, update, delete, isolate, and customize long-term memories while integrating them into a dynamic chat loop. We showed how semantic memory retrieval transforms generic assistants into context-aware systems capable of personalization and continuity across sessions. With this foundation in place, we are now equipped to extend the architecture into multi-agent systems, enterprise-grade deployments, alternative vector databases, and advanced agent frameworks, turning memory into a core capability rather than an afterthought.<\/p>\n
\n
Check out\u00a0the\u00a0<\/a>Full Implementation Code and Notebook<\/a><\/strong>.\u00a0<\/strong>Also,\u00a0feel free to follow us on\u00a0Twitter<\/mark><\/a><\/strong>\u00a0and don\u2019t forget to join our\u00a0130k+ ML SubReddit<\/a><\/strong>\u00a0and Subscribe to\u00a0our Newsletter<\/a><\/strong>. Wait! are you on telegram?\u00a0now you can join us on telegram as well.<\/a><\/strong><\/p>\n
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.?\u00a0Connect with us<\/mark><\/a><\/strong><\/p>\n
The post How to Build a Universal Long-Term Memory Layer for AI Agents Using Mem0 and OpenAI<\/a> appeared first on MarkTechPost<\/a>.<\/p>","protected":false},"excerpt":{"rendered":"
In this tutorial, we build a u…<\/p>\n","protected":false},"author":1,"featured_media":29,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-731","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/posts\/731","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=731"}],"version-history":[{"count":0,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/posts\/731\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=\/wp\/v2\/media\/29"}],"wp:attachment":[{"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=731"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=731"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/connectword.dpdns.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=731"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}