運行中 — VPS 168.231.118.131Running — VPS 168.231.118.131

你的 AI 記憶層Your AI's Memory Layer

從對話中自動提取事實、建立用戶畫像、語義搜尋 — 一個自我託管的記憶引擎

Extracts facts from conversations, builds user profiles, semantic search — a self-hosted memory engine

📊 即時數據Live Status

150
已處理 documents(ingestion 啟動以嚟)
Documents processed since ingestion
288
已匯入 documents(4 個 containers)
Imported documents (4 containers)
36
已提取 memories
Extracted memories
0
Static facts(穩定個人事實)
Static facts (stable personal facts)
Chat history ingestion 進度(288 transcripts + 6 seeds,2 concurrent)
Chat history ingestion progress (288 transcripts + 6 seeds, 2 concurrent)
Snapshot: 2026-08-15 00:21 UTC · 完成後 watchdog 會 Telegram 通知
Snapshot: 2026-08-15 00:21 UTC · watchdog will notify via Telegram when done
ContainerDocs狀態MemoriesStaticProfile (static/dynamic)
hermes103done: 41, failed: 1, queued: 61 100 2 / 100
study_lyons70done: 33, extracting: 1, failed: 3, queued: 33 100 1 / 100
prices_investigator_director113done: 40, extracting: 4, failed: 4, indexing: 2, queued: 63 100 2 / 100
security_reverse_director2done: 1, queued: 1 60 0 / 6

🧠 核心功能Core Features

🕸️ 記憶提取Memory Extraction

Memory Extraction

每個 document 由 memory agent(LLM)分析,自動抽取出「值得記住嘅事實」,唔係倒垃圾入去 — 噪音唔會變成永久記憶。

The memory agent (LLM) analyzes every document and extracts facts worth remembering — noise never becomes permanent memory.

👤 用戶畫像User Profiles

User Profiles

自動維護 static(穩定事實)+ dynamic(近期活動)兩層畫像。一次 API call ~50ms 攞晒。

Auto-maintained static + dynamic profile layers. One API call, ~50ms.

🔍 混合搜尋Hybrid Search

Hybrid Search

RAG + Memory 一條 query:知識庫文件 + 個人化上下文一齊返。

RAG + Memory in a single query: knowledge base docs + personalized context together.

自動遺忘與矛盾處理Forgetting & Contradictions

Forgetting & Contradictions

「啱啱搬咗去 SF」自動取代「住喺 NYC」;臨時事實到期自動過期。

"Just moved to SF" supersedes "lives in NYC"; temporary facts expire automatically.

📄 多模態提取Multi-modal Extraction

Multi-modal Extraction

PDF、圖片(OCR)、影片(轉錄)、code(AST-aware chunking)— 上傳就自動處理。

PDFs, images (OCR), videos (transcription), code (AST-aware chunking) — upload and it works.

🔌 ConnectorsConnectors

Connectors

Google Drive / Gmail / Notion / OneDrive / GitHub 自動同步(雲端版功能;self-hosted binary 唔包)。

Auto-sync from Drive, Gmail, Notion, OneDrive, GitHub (cloud-only; not in the self-hosted binary).

⚙️ 一個 document 點樣變成記憶How a Document Becomes Memory

1

POST /v3/documents

add 內容 + containerTag,毫秒級回傳 queued

add content + containerTag, returns queued in ms

2

Queue

ingestion queue 限速處理,搜尋永遠唔使等

rate-limited queue; search never waits

3

Chunking

內容切成語義 chunks(大 doc 可分 80+ chunks)

content split into semantic chunks

4

Embedding

本地 Xenova/bge-base-en-v1.5(768d)向量化

local Xenova/bge-base-en-v1.5 (768d) vectors

5

Memory Agent

LLM 讀全部 chunks,提取 memories(我哋用 deepseek-v4-flash)

LLM reads all chunks, extracts memories (deepseek-v4-flash)

6

Memory Graph

facts 入 graph:版本化、矛盾解決、過期

facts enter the graph: versioned, contradictions resolved

7

Search / Profile

語義搜尋 + 畫像即時可用(~50ms)

semantic search + profile live (~50ms)

🏗️ 我哋嘅部署架構Our Deployment Architecture

Mac (Hermes 4 profiles) │ memory.provider = supermemory ▼ 127.0.0.1:6767 ←─ autossh SSH tunnel(launchd 常駐、斷線自動重連) │ SSH 加密,只轉發 6767 ▼ VPS 168.231.118.131 (Ubuntu 24.04) ├─ supermemory-server v0.0.7-rc.2(PINNED,sha256 f56e7abc…d9a8c) │ ├─ data: /var/lib/supermemory(encrypted local storage) │ ├─ embeddings: 本地 Xenova/bge-base-en-v1.5(768d,免 API key) │ └─ 綁 0.0.0.0:6767 + iptables 只准 loopback(外網封死) └─ LLM (memory extraction): opencode-go → https://opencode.ai/zen/go/v1 model: deepseek-v4-flash(同本機 Hermes 共用配額)
💡 containers 各自隔離:hermes / study_lyons / prices_investigator_director / security_reverse_director — 每個 profile 嘅記憶互不相干,但共用同一個 server 同埋同一個 LLM。
💡 Containers are isolated per profile: hermes / study_lyons / prices_investigator_director / security_reverse_director — each profile's memory is separate, sharing one server and one LLM.

🏆 BenchmarksBenchmarks

Benchmark結果Result
LongMemEval(長期記憶)#1 — 95% Recall@15,上下文縮減 99.4%
LoCoMo(多跳/時序推理)#1
ConvoMem(個人化)#1
Memory vs RAGRAG 只檢索 document chunks;Memory 提取並追蹤「關於用戶嘅事實」— 兩者一齊行(hybrid)RAG retrieves chunks; Memory extracts facts about users — both run together (hybrid)

🔌 API 速覽(localhost:6767)API at a Glance (localhost:6767)

EndpointMethod用途Purpose
/v3/documentsPOSTadd 內容(text/URL/HTML)add content
/v4/searchPOST混合搜尋(hybrid/memories/documents)hybrid search
/v4/profilePOST用戶畫像 + 相關記憶user profile + relevant memories
/v4/memories/listPOST列出 memories(isStatic 標記)list memories
/v4/openapiGET完整 OpenAPI spec(602KB)full OpenAPI spec
/GETlocal quickstart 頁local quickstart page