内存管理

存储用户偏好和反馈，让你的 LLM 在不同会话间记住这些信息。按用户和团队进行范围界定，并内置访问控制。

要求： LiteLLM v1.83.10+ 并连接 PostgreSQL。无需更改配置。

创建

curl

curl -X POST "https://:4000/v1/memory" \
  -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" \
  -d '{
    "key": "user:preferences",
    "value": "Prefers concise responses. Timezone: PST.",
    "metadata": {"version": 1}
  }'

Python

import httpx

client = httpx.Client(
    base_url="https://:4000",
    headers={"Authorization": "Bearer sk-1234"},
)

client.post("/v1/memory", json={
    "key": "user:preferences",
    "value": "Prefers concise responses. Timezone: PST.",
    "metadata": {"version": 1},
})

读取

curl "https://:4000/v1/memory/user:preferences" \
  -H "Authorization: Bearer sk-1234"

更新

curl -X PUT "https://:4000/v1/memory/user:preferences" \
  -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" \
  -d '{"value": "Prefers concise responses. Timezone: EST."}'

列表

# All entries
curl "https://:4000/v1/memory" \
  -H "Authorization: Bearer sk-1234"

# By prefix
curl "https://:4000/v1/memory?key_prefix=user:" \
  -H "Authorization: Bearer sk-1234"

删除

curl -X DELETE "https://:4000/v1/memory/user:preferences" \
  -H "Authorization: Bearer sk-1234"

访问控制

基于 API 密钥自动进行范围界定。

角色	读取	写入
用户	自身 + 团队条目	仅限自身条目
团队管理员	自身 + 团队条目	自身 + 团队条目
代理（Proxy）管理员	全部	全部

键命名

键（Key）是全局唯一的。使用前缀进行命名空间划分和查询

user:preferences           → per-user settings
team:playbook:onboarding   → shared team resources
agent:memory:scratchpad    → agent working memory

示例：Slack 机器人中的按用户内存

按 Slack 工作区和用户划分内存，从而隔离每个人的偏好。

键格式： slack:{team_id}:{user_id}

import httpx

LITELLM_BASE = "https://:4000"
LITELLM_KEY = "sk-1234"

def memory_key(team_id: str, user_id: str) -> str:
    return f"slack:{team_id}:{user_id}"

async def get_preferences(team_id: str, user_id: str) -> str:
    """Read saved preferences. Returns "" if none exist."""
    key = memory_key(team_id, user_id)
    async with httpx.AsyncClient() as client:
        r = await client.get(
            f"{LITELLM_BASE}/v1/memory/{key}",
            headers={"Authorization": f"Bearer {LITELLM_KEY}"},
        )
    if r.status_code == 404:
        return ""
    return r.json().get("value", "")

async def save_preference(team_id: str, user_id: str, note: str):
    """Append a preference. PUT upserts — creates or updates."""
    key = memory_key(team_id, user_id)
    existing = await get_preferences(team_id, user_id)

    # Store as bullet list
    bullets = [b for b in existing.split("\n") if b.strip()]
    bullets.append(f"- {note}")
    
    async with httpx.AsyncClient() as client:
        await client.put(
            f"{LITELLM_BASE}/v1/memory/{key}",
            headers={"Authorization": f"Bearer {LITELLM_KEY}"},
            json={"value": "\n".join(bullets)},
        )

在每一轮对话中注入到你的系统提示词（System Prompt）中

prefs = await get_preferences(team_id, user_id)

messages = [
    {"role": "system", "content": f"""You are a helpful assistant.

SAVED USER PREFERENCES:
{prefs}

Follow these unless the current message contradicts them."""},
    {"role": "user", "content": user_message},
]

查询工作区的所有偏好设置

curl "https://:4000/v1/memory?key_prefix=slack:T024BE7LD:" \
  -H "Authorization: Bearer sk-1234"

元数据

为条目附加任何 JSON 数据

{
  "key": "agent:findings",
  "value": "Q1 API usage up 15%...",
  "metadata": {"tags": ["research"], "confidence": 0.92}
}

API 参考

完整的请求/响应架构、参数和错误代码：/memory 端点参考。