https://www.zhihu.com/question/2039044882887616442 ，已经有人研究了怎么做。常用的软件有 mem0 ，Hindsight。
我的做法
nas中布署了 hindsight-slim 的服务。使用阿里百炼的 text-embedding-v4 做向量化。然后 hindsight 接的 deepseek-flash做后端大模型。实测可用，速度和效果暂时都还觉得不错。
分享一下配置：
version: "3.8"
services:
db:
image: pgvector/pgvector:pg16
container_name: hindsight-db
environment:
POSTGRES_USER: hindsight
POSTGRES_PASSWORD: mmm
POSTGRES_DB: hindsight
volumes:
- /share/Docker/ai-memory/postgres-data:/var/lib/postgresql/data
restart: unless-stopped
deploy:
resources:
limits:
memory: 512M
cpus: '0.5'
hindsight:
image: ghcr.io/vectorize-io/hindsight:latest-slim
container_name: hindsight-app
ports:
- "18888:8888"
- "9999:9999"
volumes:
- /share/Docker/ai-memory/hindsight-data:/home/hindsight/.pg0
environment:
TZ: Asia/Shanghai
HINDSIGHT_DB_HOST: db
HINDSIGHT_DB_PORT: 5432
HINDSIGHT_DB_NAME: hindsight
HINDSIGHT_DB_USER: hindsight
HINDSIGHT_DB_PASSWORD: mmm
HINDSIGHT_API_EMBEDDINGS_PROVIDER: openai
HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL: https://ws-x0pz7lt9imoytqek.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
HINDSIGHT_API_EMBEDDINGS_OPENAI_BATCH_SIZE: "10"
HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: xxx
OPENAI_API_KEY: xxx
HINDSIGHT_API_OPENAI_API_KEY: xxx
HINDSIGHT_API_KEY: xxx
HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL: text-embedding-v4
HINDSIGHT_API_LLM_PROVIDER: openai
HINDSIGHT_API_LLM_BASE_URL: https://api.deepseek.com/v1
HINDSIGHT_API_LLM_API_KEY: xxx
HINDSIGHT_API_LLM_MODEL: deepseek-chat
HINDSIGHT_API_LLM_TIMEOUT: 60
HINDSIGHT_API_RERANKER_PROVIDER: rrf
HINDSIGHT_CONSOLIDATE_INTERVAL: 86400
HINDSIGHT_API_RECALL_MAX_CONCURRENT: 4
HINDSIGHT_API_WORKER_REFRESH_MENTAL_MODEL_MAX_SLOTS: 1
HINDSIGHT_API_WORKER_RETAIN_MAX_SLOTS: 1
HINDSIGHT_API_DB_POOL_MAX_SIZE: 50
HINDSIGHT_API_TENANT_EXTENSION: hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
HINDSIGHT_API_TENANT_API_KEY: yyy
HINDSIGHT_CP_ACCESS_KEY: yyy
HINDSIGHT_API_MCP_AUTH_TOKEN: yyy
deploy:
resources:
limits:
memory: 1.8G
cpus: '1.0'
depends_on:
- db
restart: unless-stopped