SAME - Stateless Agent Memory Engine

io.github.sgx-labs由 GitHub 帳號 sgx-labs 發布v0.12.1更新於 Oct 25, 2026

Persistent memory for AI coding agents. Local-first SQLite + vector search. 17 MCP tools.

發布方自己的服務:Memory.fit

已驗證STDIO僅桌面知識與記憶開發者工具

概覽

AI 產生的概覽

SAME 是一個本機優先的記憶引擎,讓 AI 程式助理對一組 Markdown 筆記擁有持久、可搜尋的記憶。

可以做什麼
搜尋與讀取筆記:語意搜尋、帶篩選的搜尋、跨多個 vault 的聯合搜尋、依路徑讀取完整筆記、尋找相似筆記、最近活動,以及包含置頂筆記和交接的 session 上下文。寫入記憶:建立或更新筆記、記錄結構化決策、撰寫 session 交接與改進事項。維護索引:重建索引、查看索引統計、合併相關筆記、產生簡報和 vault 健康報告。管理封鎖:將筆記從搜尋結果中隱藏、還原,以及列出已封鎖的筆記。
適用情境
當你希望 AI 程式工具在多次 session 之間,以及在 Claude Code、Cursor、Windsurf 等不同編輯器之間保留決策、上下文和交接,並以自己的 Markdown 筆記作為記憶儲存時適用。
執行需求
透過 npm 或安裝腳本作為本機程序執行;必須以 VAULT_PATH 環境變數指向你的 Markdown 筆記目錄。嵌入是選用的:關鍵字搜尋不需要任何外部相依,語意搜尋則需要本機 Ollama 實例或 OpenAI 或相容 OpenAI 的嵌入服務。
安裝前請注意
此伺服器會讀寫筆記目錄中的檔案,包括建立和更新筆記、決策與交接。若設定雲端嵌入服務,筆記內容會被傳送到該服務進行嵌入。不要把機密放在會被索引的目錄中;README 建議使用一個從不被索引的私有目錄。選用的 HTTP 模式會加入 Bearer 權杖驗證。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 SAME - Stateless Agent Memory Engine,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

SAME — Persistent Memory for AI Coding Agents

[License: BSL 1.1] [Go] [Latest Release] [GitHub Stars] [MCP Tools] [Discord]

Your AI forgets everything between sessions. SAME fixes that.

SAME gives every AI coding tool persistent memory. Claude Code, Cursor, Windsurf, Codex CLI, Gemini CLI — one memory layer that works everywhere. It indexes your markdown notes, surfaces relevant context automatically, and records decisions and handoffs so your AI picks up where it left off.

One binary. Fully local. No cloud. No telemetry. Mac, Linux, Windows, Raspberry Pi.

Install

bash
# macOS / Linuxcurl -fsSL https://statelessagent.com/install.sh | bash
# Windows (PowerShell)irm https://statelessagent.com/install.ps1 | iex

Or via npm (all platforms): npm install -g @sgx-labs/same

Installed via npm? Update with npx same@latest or npm update -g @sgx-labs/same.

See It Work

bash
same demo
Indexing 5 sample notes...Searching: "authentication decision"
  1. decisions/auth-strategy.md (score: 0.94)     "We chose JWT with refresh tokens for..."
  2. notes/api-security.md (score: 0.87)     "Auth middleware validates tokens at..."
Asking: "what did we decide about authentication?"
  Based on your notes, you decided to use JWT with refresh  tokens (decisions/auth-strategy.md). The API middleware  validates tokens at the gateway level (notes/api-security.md).
No accounts. No API keys. Everything runs locally.

Quickstart

bash
# 1. Point SAME at your projectcd ~/my-project && same init
# 2. Test searchsame search "authentication decision"
# 3. Done. Your AI now has memory.# Start Claude Code, Cursor, or any MCP client.

same init sets up hooks and MCP tools automatically. Your AI gets relevant context on every session start.

Key Features

  • Your AI remembers everything -- Decisions, handoffs, and context survive across sessions. Close your terminal, switch projects, come back tomorrow. Nothing gets lost.

  • Memory integrity -- Tracks provenance (where notes came from), detects when source files change, and flags stale knowledge. Stale notes rank lower in search automatically. same health shows trust state across your vault.

  • Dual-layer memory -- Extracts atomic facts from your notes via LLM. Facts are independently searchable and boost source notes in search results. The right answer surfaces even when the fact is buried in an unrelated conversation.

  • Streamable HTTP transport -- same web --mcp enables an HTTP MCP endpoint with Bearer token auth. Connect from Open WebUI, LobeChat, or any HTTP MCP client — no stdio required.

  • Works with your tools -- 19 MCP tools for Claude Code, Cursor, Windsurf, or any MCP client. Search, save decisions, create handoffs without leaving your editor.

  • Safe for teams -- Multiple AI agents on the same codebase won't step on each other. File claims, push protection, and attribution built in.

  • Instant expertise -- 17 pre-built knowledge vaults with 870+ curated notes. One command to install. Your AI gets domain knowledge in seconds.

  • Connected knowledge -- See how decisions, files, and notes relate to each other. Ask "what depends on this?" and get real answers. Powered by SQLite.

Security & Teams

SAME includes built-in PII scanning and push protection:

  • PII scanning -- Pre-commit hooks detect emails, API keys, secrets, and personal data before they reach git. Configurable blocklists with false-positive review workflow.
  • Push protection -- Multi-agent file claims prevent AI agents from overwriting each other's work. Advisory locks with attribution.
  • Audit logging -- Every guard scan, every allow decision, every override is logged.
  • Privacy tiers -- _PRIVATE/ is never indexed. research/ is indexed but never committed. Your notes, your rules.
bash
same guard settings set push-protect on    # enable push protectionsame guard scan                            # run PII scan manually

How It Works

Your Notes (.md)  -->  Embeddings  -->  SQLite  -->  Your AI Tool                       (local or        (search      (Claude Code,                        cloud)           + rank)      Cursor, etc.)

Your markdown notes get embedded and stored in SQLite. When your AI starts a session, SAME surfaces relevant context via hooks or MCP. Decisions get extracted. Handoffs get generated. The next session picks up where the last one stopped.

No Ollama? No problem. SAME runs with zero external dependencies using keyword search (SQLite FTS5). Add Ollama later for semantic search -- same reindex upgrades instantly.

Why SAME

Without SAMEWith SAME
Re-explain everything each sessionAI picks up where you left off
"Didn't we decide to use JWT?"Decision surfaces automatically
"Is this note still accurate?"Trust state flags stale knowledge
Close terminal = context lostHandoff recovers the session
Copy-paste notes into chatsame ask with source citations
Context compacted mid-taskPinned notes survive compaction

The Numbers

MetricValue
Recall@5100% keyword, 84% semantic on internal eval (68 cases). Held-out: 90% Recall@5 on 30 blind cases (see eval/METHODOLOGY.md)
MRR0.65 keyword, 0.62 semantic
Prompt overhead<200ms
Binary size~14MB
Setup timeUnder 2 minutes

Add to Your AI Tool

Claude Code (recommended)

bash
same init    # installs 6 hooks + MCP automatically

Cursor / Windsurf / Any MCP Client

Add to your MCP config (.mcp.json, Cursor settings, etc.):

json
{  "mcpServers": {    "same": {      "command": "npx",      "args": ["-y", "@sgx-labs/same", "mcp", "--vault", "/path/to/your/notes"]    }  }}

19 MCP tools available instantly. Works without Ollama (keyword fallback).

Switch between Claude Code and Cursor without losing context. Your memory travels with you.

Tool Compatibility

Claude Code gets full automatic handoffs via hooks. Cursor, Windsurf, Codex CLI, Gemini CLI get full MCP tool access (search, save, decisions, graph) but handoffs need to be triggered manually. We're working on automatic handoff support for more editors.

MCP Server

ToolWhat it does
search_notesSemantic search across your knowledge base
search_notes_filteredSearch with domain/tag/agent filters
search_across_vaultsFederated search across multiple vaults
get_noteRead full note content by path
find_similar_notesDiscover related notes
get_session_contextPinned notes + latest handoff + git state
recent_activityRecently modified notes
save_noteCreate or update a note
save_decisionLog a structured project decision
create_handoffWrite a session handoff
reindexRe-scan and re-index the vault
index_statsIndex health and statistics
mem_consolidateConsolidate related notes via LLM
mem_briefGenerate orientation briefing
mem_healthVault health with trust analysis
mem_forgetSuppress a note from search results
mem_restoreUndo mem_forget (unsuppress a note)
mem_list_suppressedList suppressed notes
save_kaizenLog improvement items with provenance

SeedVaults

Pre-built knowledge vaults. One command to install.

bash
same seed list                              # browse available seedssame seed install claude-code-power-user    # install one
SeedNotesWhat you get
same-getting-started18Learn SAME itself — the universal on-ramp
claude-code-power-user50Claude Code workflows and operational patterns
ai-agent-architecture56Agent design, orchestration, memory strategies
api-design-patterns56REST, GraphQL, auth, rate limiting, and more
typescript-fullstack-patterns55Full-stack TypeScript patterns and best practices
engineering-management-playbook59Engineering leadership and team management
personal-productivity-os117GTD, time blocking, habit systems
security-audit-framework61Security review checklists and frameworks

Plus 9 more. Browse all 17 seeds on GitHub.

Privacy

All data stays on your machine. SAME creates a three-tier privacy structure:

DirectoryIndexedCommittedUse for
Your notesYesYour choiceDocs, decisions, research
_PRIVATE/NoNoAPI keys, credentials
research/YesNoStrategy, analysis

No telemetry. No cloud. Path traversal blocked. Config files written with owner-only permissions.

More

Full CLI Reference
CommandDescription
same initSet up SAME for your project
same demoSee SAME in action with sample notes
same tutorial7 hands-on lessons
same ask <question>Ask a question, get cited answers
same search <query>Search your notes
same search --all <query>Search across all vaults
same statusSee what SAME is tracking
same doctorRun diagnostic checks
same claim <path> --agent <name>Advisory file ownership for multi-agent
same pin <path>Always include a note in sessions
same graph statsKnowledge graph diagnostics
same webLocal web dashboard
same seed listBrowse available seed vaults
same seed install <name>Install a seed vault
same vault list|add|remove|defaultManage multiple vaults
same guard settings set push-protect onEnable push protection
same consolidateMerge related notes into knowledge summaries
same briefAI-generated orientation briefing
same healthVault health score with trust/provenance analysis
same staleList all stale notes in your vault
same search --trust staleFilter search by trust state
same search --type decisionFilter search by content type
same ignoreView/manage .sameignore patterns
same factsView, search, and manage extracted facts
same config set <key> <value>Set config values from CLI
same brief --no-llmStructured briefing without LLM
same tipsBest practices for vault hygiene and security
same reindex [--force]Rebuild search index
same repairBack up and rebuild database
same updateUpdate to latest version
same completion [bash|zsh|fish]Shell completions
Configuration

SAME uses .same/config.toml, generated by same init:

toml
[vault]path = "/home/user/notes"handoff_dir = "sessions"decision_log = "decisions.md"
[embedding]provider = "ollama"           # "ollama", "openai", "openai-compatible", or "none"model = "nomic-embed-text"
[memory]max_token_budget = 800max_results = 2

Supported embedding models: nomic-embed-text (default), snowflake-arctic-embed2, mxbai-embed-large, all-minilm, text-embedding-3-small (OpenAI), and more.

Configuration priority (highest wins): CLI flags > Environment variables > Config file > Defaults

More Install Options
bash
# Dockergit clone --depth 1 https://github.com/sgx-labs/statelessagent.gitcd statelessagent && docker build -t same .
# Build from source (requires Go 1.25+)git clone --depth 1 https://github.com/sgx-labs/statelessagent.gitcd statelessagent && make install
Troubleshooting

Start with same doctor -- it runs 20+ checks and tells you what's wrong.

"No vault found" -- Run same init from inside your notes folder, or set VAULT_PATH=/path/to/notes.

"Ollama not responding" -- SAME falls back to keyword search automatically. Test with curl http://localhost:11434/api/tags.

Hooks not firing -- Run same setup hooks to reinstall. Verify with same status.

Database issues -- Run same repair to back up and rebuild.

SAME vs. Alternatives
SAMEmem0LettaCLAUDE.md
Setup1 commandpip + configpip or DockerEdit file
Runtime depsNonePython + vector DBPython + SQLAlchemyNone
OfflineFullNot defaultWith local modelsYes
Cloud requiredNoDefault yesNoNo
TelemetryNoneDefault ONYesNone
MCP tools199Client onlyNo
Memory integrityProvenance + trustNoNoNo
Knowledge graphBuilt-inRequires Neo4jNoNo
Cross-tool memoryYesAPI onlyNoClaude only
Runs on PiYes (~14MB)NoNoYes
Eval Methodology

Internal evaluation on 105 tuning cases. Held-out validation: 93.3% Recall@5 on 30 blind test cases (see eval/METHODOLOGY.md).

MetricValueDataset
Recall@5 (keyword)100%Internal (68 cases)
Recall@5 (semantic)84%Internal (68 cases)
MRR (keyword)0.65Internal (68 cases)
Recall@590%Held-out (30 blind cases)

All evaluation uses synthetic vault data. No user data used.

Links

Contributing

Contributions welcome. Open an issue or start a discussion.

bash
git clone https://github.com/sgx-labs/statelessagent.gitcd statelessagentmake build && make test

See SECURITY.md for security-related reports.

Support

Buy me a coffee | GitHub Sponsors

License

BSL 1.1. Free for personal, educational, hobby, research, and evaluation use. Converts to Apache 2.0 on 2030-02-02. See LICENSE.


[圖片]

來源:README.md,提交 85e7ff2

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

1
  1. v0.12.1最新Sep 16, 2026

評分

已滿足適用檢查項的 40%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。

其他 Memory 整合