🧠 Mem055k Stars AI Agent LLM：

🧠 Mem0：55k Stars 的 AI Agent 记忆层，让你的 LLM 不再金鱼记忆

项目地址：https://github.com/mem0ai/mem0 | ⭐ 55.3k Stars | 🛠 Python | 作者：mem0ai

老实说，现在很多 AI Agent 看着挺聪明，但聊几句就发现——它根本不记得你上一轮说了啥。每次对话都是全新的开始，跟金鱼似的。

Mem0（"mem-zero"）就是来解决这个问题的。它给 AI Agent 装了一个长期记忆层，能记住用户的偏好、习惯，甚至跨 session 持续学习。最新的 v3 算法在 LoCoMo 基准上拿了 91.6 分（比上代 +20 分），LongMemEval 93.4 分，关键是每次检索只用 ~7K tokens、1 秒延迟。

一行命令装上去

pip install mem0ai

如果要 BM25 关键词匹配和实体抽取，加个 NLP 依赖：

pip install mem0ai[nlp]
python -m spacy download en_core_web_sm

想直接在终端里管理记忆也行：

npm install -g @mem0/cli
# 或者 pip install mem0-cli

mem0 init
mem0 add "Prefers dark mode and vim keybindings" --user-id alice
mem0 search "What does Alice prefer?" --user-id alice

给 LLM 装上记忆，代码就这几行

from openai import OpenAI
from mem0 import Memory

client = OpenAI()
memory = Memory()

def chat_with_memories(message: str, user_id: str = "default_user") -> str:
    relevant = memory.search(query=message, filters={"user_id": user_id}, top_k=3)
    memories_str = "\n".join(f"- {entry['memory']}" for entry in relevant["results"])

    system_prompt = f"You are a helpful AI. Answer based on query and memories.\nUser Memories:\n{memories_str}"
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "system", "content": system_prompt},
                   {"role": "user", "content": message}]
    )
    answer = response.choices[0].message.content

    memory.add(
        [{"role": "user", "content": message},
         {"role": "assistant", "content": answer}],
        user_id=user_id
    )
    return answer

最骚的操作是——每次对话结束时 memory.add() 自动把这次聊的内容存进去，下次 memory.search() 就能查到。Agent 越用越懂你。

自部署也简单

cd server && docker compose up -d
# 浏览器打开 http://localhost:3000

用 make bootstrap 一步到位：启动服务、创建管理员、生成 API key。

v3 新算法，值得升级

| 基准 | 旧版 | v3 | Tokens | 延迟 |

|------|------|----|--------|------|

| LoCoMo | 71.4 | 91.6 | 7.0K | 0.88s |

| LongMemEval | 67.8 | 93.4 | 6.8K | 1.09s |

| BEAM (1M) | — | 64.1 | 6.7K | 1.00s |

核心改进：单次 ADD-only 抽取（不更新不删除），实体链接跨记忆关联，语义+BM25+实体三重检索融合。

踩坑提醒

总结：

2. pip install mem0ai + 几行代码就能集成

3. v3 算法在长上下文记忆 recall 上比旧版提升了 20+ 分

4. 支持本地部署、CLI、以及 Claude Code/Codex 的 skills 集成

5. Apache 2.0 开源，YC S24 团队出品

English Version

🧠 Mem0: 55k Stars Memory Layer for AI Agents — Your LLM Doesn't Have to Be a Goldfish

Repo: https://github.com/mem0ai/mem0 | ⭐ 55.3k Stars | 🛠 Python

Let's be honest — most AI agents today are smart until you realize they forgot everything you said two messages ago. Each conversation is a fresh start, like talking to a goldfish.

Mem0 (pronounced "mem-zero") is a universal memory layer that gives AI agents long-term recall. It remembers user preferences, adapts over time, and learns across sessions. The v3 algorithm scores 91.6 on LoCoMo (+20 over the previous version), 93.4 on LongMemEval, while using only ~7K tokens and ~1s latency per retrieval.

One-line install

pip install mem0ai

With NLP support for BM25 keyword matching and entity extraction:

pip install mem0ai[nlp]
python -m spacy download en_core_web_sm

Or manage memories from your terminal:

npm install -g @mem0/cli
mem0 init
mem0 add "Prefers dark mode and vim keybindings" --user-id alice
mem0 search "What does Alice prefer?" --user-id alice

Memory in 15 lines of Python

from openai import OpenAI
from mem0 import Memory

client = OpenAI()
memory = Memory()

def chat_with_memories(message: str, user_id: str = "default_user") -> str:
    relevant = memory.search(query=message, filters={"user_id": user_id}, top_k=3)
    memories_str = "\n".join(f"- {entry['memory']}" for entry in relevant["results"])
    # ... inject memories into prompt, call LLM, then store new memories
    return answer

The trick: memory.add() automatically stores each conversation turn after the LLM responds. Next time memory.search() retrieves what's relevant. The more you use it, the smarter it gets.

Self-hosting

cd server && docker compose up -d
# Open http://localhost:3000

Or make bootstrap for one-command setup with admin creation and API key generation.

Key takeaways

2. pip install mem0ai + a few lines of Python and you're integrated

3. v3 algorithm delivers 20+ point improvements over previous version on long-context benchmarks

4. Supports local deployment, CLI, and skills for Claude Code/Codex

5. Apache 2.0 licensed, YC S24 team