能力标签

🔌 MCP 🤖 Agent 🔄 工作流 💻 CLI 🔗 REST API 🧬 Embedding 🧠 Claude ✨ GPT ⌨️ Cursor 🖥 本地 LLM

🔌

MCP工具

非线性内存OS

基于 TypeScript · 让 AI 助手直接操作你的系统与工具

英文名：nlm-memory

⭐ 6 Stars 💻 TypeScript 📄 Apache-2.0 🏷 AI 7.5分

7.5AI 综合评分

aiaiderclaude-codecodexcursortypescript

⬇ 下载源码 ZIP ⚙️ 配置说明

✦ AI Skill Hub 推荐

非线性内存OS 是 AI Skill Hub 本期精选MCP工具之一。综合评分 7.5 分，整体质量较高。我们推荐使用将其纳入你的 AI 工具库，帮助提升工作效率。

📚 深度解析

非线性内存OS 是一款基于 MCP（Model Context Protocol）标准协议的 AI 工具扩展。MCP 协议由 Anthropic 开发并开源，旨在建立 AI 模型与外部工具之间的标准化通信接口，目前已被 Claude Desktop、Claude Code、Cursor 等主流 AI 工具采纳。

通过安装非线性内存OS，你的 AI 助手将获得额外的工具调用能力，可以用自然语言直接操控该工具的功能，无需学习复杂的命令行语法。MCP 工具的核心价值在于"一次配置，永久增强"——配置完成后，每次与 AI 对话时都可以无缝调用这些工具。

在技术实现上，MCP 工具通过标准的 JSON-RPC 协议与 AI 客户端通信，工具的功能以"工具列表"的形式暴露给 AI 模型，AI 可以按需调用。非线性内存OS 提供了结构化的工具调用接口，使 AI 模型能够精确地理解和使用每个功能点，显著降低 AI 在工具使用上的错误率。

与传统的 API 集成相比，MCP 工具的优势在于无需编写代码——用户只需在配置文件中添加几行 JSON，即可让 AI 获得全新能力。AI Skill Hub 将非线性内存OS 评为 AI 评分 7.5 分，属于同类工具中的优质选择。

📋 工具概览

非线性内存OS 是一款遵循 MCP（Model Context Protocol）标准协议的 AI 工具扩展。通过 MCP 协议，它可以让 Claude、Cursor 等主流 AI 客户端直接访问和操作外部工具、数据源和服务，实现 AI 能力的无缝扩展。无论是文件操作、数据库查询还是 API 调用，都可以通过自然语言在 AI 对话中直接触发，极大提升生产效率。

GitHub Stars

⭐ 6

开发语言

TypeScript

支持平台

Windows / macOS / Linux

维护状态

轻量级项目，按需更新

开源协议

Apache-2.0

AI 综合评分

7.5 分

工具类型

MCP工具

Forks

—

📖 中文文档

以下内容由 AI Skill Hub 根据项目信息自动整理，如需查看完整原始文档请访问底部「原始来源」。

📌 核心特色

通过标准 MCP 协议与 Claude、Cursor 等主流 AI 客户端深度集成
提供结构化工具调用接口，显著降低 AI 集成复杂度
支持 Claude Desktop 和 Claude Code 无缝接入，开箱即用
可与其他 MCP 工具组合叠加，构建完整 AI 工作站
轻量无侵入设计，不影响现有系统架构

🎯 主要使用场景

在 Claude Desktop 对话中直接调用本地工具，实现 AI 与系统的深度联动
通过自然语言驱动复杂的多步骤自动化任务，代替繁琐手动操作
将多个 MCP 工具组合使用，构建个人专属 AI 工作站

以下安装命令基于项目开发语言和类型自动生成，实际以官方 README 为准。

安装命令

# 方式一：通过 Claude Code CLI 一键安装
claude skill install https://github.com/pbmagnet4/nlm-memory

# 方式二：手动配置 claude_desktop_config.json
{
  "mcpServers": {
    "-----os": {
      "command": "npx",
      "args": ["-y", "nlm-memory"]
    }
  }
}

# 配置文件位置
# macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
# Windows: %APPDATA%/Claude/claude_desktop_config.json

📋 安装步骤说明

确认已安装 Node.js（v18 或以上版本）
打开 Claude Desktop 或 Claude Code 的 MCP 配置文件
按「交给 Agent 安装 → Claude Desktop」标签中的 JSON 配置填入 mcpServers 字段
保存配置文件并重启 Claude 客户端
重启后，在对话中即可使用本工具

以下用法示例由 AI Skill Hub 整理，涵盖最常见的使用场景。

常用命令 / 代码示例

# 安装后在 Claude 对话中直接使用
# 示例：
用户: 请帮我用 非线性内存OS 执行以下任务...
Claude: [自动调用 非线性内存OS MCP 工具处理请求]

# 查看可用工具列表
# 在 Claude 中输入："列出所有可用的 MCP 工具"

以下配置示例基于典型使用场景生成，具体参数请参照官方文档调整。

配置示例

// claude_desktop_config.json 配置示例
{
  "mcpServers": {
    "_____os": {
      "command": "npx",
      "args": ["-y", "nlm-memory"],
      "env": {
        // "API_KEY": "your-api-key-here"
      }
    }
  }
}

// 保存后重启 Claude Desktop 生效

📑 README 深度解析真实文档完整度 81/100 查看 GitHub 原文 →

以下内容由系统直接从 GitHub README 解析整理，保留代码块、表格与列表结构。

简介

<a href="#install">Install</a> · <a href="#quick-start">Quick Start</a> · <a href="#runtimes">Runtimes</a> · <a href="#how-recall-works">How recall works</a> · <a href="#agent-self-improvement-signals">Signals</a> · <a href="#mcp-tools">MCP</a> · <a href="#rest-api">REST API</a> · <a href="#daily-digest">Digest</a> · <a href="#configuration">Config</a> · <a href="#security">Security</a> · <a href="#vs-alternatives">vs Alternatives</a>

---

nlm-memory is a local-first memory layer for AI coding agents. It indexes every session from Claude Code, Codex, OpenCode, Cursor, Windsurf, Hermes, Aider, and pi into a single searchable store on your machine. Three properties no other memory layer ships together:

Cross-runtime reach. One index, every adapter.
Editable timeline. Sessions can be superseded by newer ones; entities can be retired. Patch history retroactively — no other tool lets you do this. See docs/supersedence.md.
97.2% R@5 baseline. On a 14-month corpus, keyword recall surfaces the right session in the top 5 on 97.2% of evaluator queries. No fine-tuning. The labels were generated by DeepSeek V4 — the retrieval algorithm is the same code path you'll run, but expect a lower number with a smaller local classifier. See docs/methodology-recall-baseline.md.

Everything stays on your machine by default. No telemetry, no account. The classifier defaults to local (Ollama); if you opt into a cloud classifier (DeepSeek, OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible endpoint), session transcripts are sent to that provider — see Security for the exact data-flow.

---

What's inside the UI

Open http://localhost:3940/ui after the daemon starts.

Page	What it shows
Live	Sessions being written in real time, recent reads, recent decisions
Pulse	System health — coherence, runtimes, stale entities, recent sessions
River	Full session timeline with density controls + superseded-lane visualization
Thread	Per-entity conversation history with runtime filters and ←/→ navigation
Search	Keyword, semantic, or hybrid recall with match snippets and field-origin tags
Recall	Adoption telemetry — useful_hit_rate, source breakdown, query log
Settings	Sources, providers, classifier, data backup/restore

---

Install

npm install -g nlm-memory
nlm setup

nlm setup is the interactive first-run wizard. It picks your classifier + model, wires the runtimes you actually use, generates an NLM_MCP_TOKEN, hardens permissions on ~/.nlm/, and installs the daemon supervisor for your platform. For headless environments (servers, CI, agents), use nlm start directly after initial setup — it boots the daemon without interactive prompts.

Platform	Daemon	Notes
macOS	LaunchAgent at `~/Library/LaunchAgents/com.github.pbmagnet4.nlm-memory.plist`	Auto-starts on login
Linux	systemd user unit at `~/.config/systemd/user/nlm.service`	Headless servers: `loginctl enable-linger $USER` so the daemon survives logout
Windows	Manual `nlm start` for now	Hook + MCP install paths are platform-aware; supervisor lands next release

Stop or remove: nlm uninstall.

---

Quick Start

After nlm setup finishes, open http://localhost:3940/ui — the daemon is running. (If you changed NLM_PORT, substitute your port in all examples below.) A 30-second sanity check:

nlm recall "what was that pgvector decision"   # one-shot search from the shell
nlm digest                                      # yesterday's activity at a glance
nlm --version

---

Configuration and privacy

Var	Default	What
`NLM_SIGNALS_ENABLED`	`1` (on)	Set to `0` to disable signal ingest entirely
`NLM_SIGNAL_RETENTION_DAYS`	`90`	Raw signals older than this are pruned

Signals are local-only. They are stamped with a per-install ID from ~/.nlm/install-id and never leave the machine.

Configuration

Environment variables

Var	Default	What
`NLM_PORT`	`3940`	Daemon bind port (loopback only)
`NLM_DB_PATH`	`~/.nlm/canonical.sqlite`	SQLite canonical store location
`NLM_HOOK_MODE`	`live`	`live` injects pointer block; `shadow` logs without injecting
`NLM_HOOK_LOG`	`~/.nlm/hook-log.jsonl`	Hook fire log; powers digest's liveness alert
`NLM_USEFUL_HIT_LOG`	`~/.nlm/useful-hit-log.jsonl`	Citation/useful-hit ledger
`NLM_QUERY_LOG`	`~/.nlm/query-log.jsonl`	Recall query telemetry
`NLM_CITATION_LOG`	`~/.nlm/citation-log.jsonl`	Stop-hook citation events
`NLM_MISS_LOG`	`~/.nlm/miss-log.jsonl`	Stop-hook miss events — sessions the agent explicitly fetched via `get_session`/`cite_session` that the hook never surfaced. Reviewed via `nlm misses`.
`NLM_MISS_LOG_ENABLED`	`true`	Set to `0` to disable miss-log emission entirely.
`NLM_MCP_TOKEN`	auto-generated	256-bit bearer for `/api/*` (non-browser) and `/mcp`
`NLM_MCP_CONFIG`	`~/.mcp.json`	Path the `connect`/`disconnect` commands modify
`NLM_CLASSIFIER`	`ollama`	`ollama` (local, default), `deepseek`, `openai`, `anthropic`, `openrouter`, or `openai-compatible`
`NLM_CLASSIFIER_MODEL`	`qwen3:4b-instruct-2507-q4_K_M`	Model id for the chosen provider. See [classifier bench](reports/classifier-comparison/2026-06-02-deepseek-v4-vs-qwen3.md) for why this is the recommended local default.
`NLM_OLLAMA_URL`	`http://localhost:11434`	Override Ollama endpoint
`NLM_ADAPTERS`	all	Comma-separated allowlist of adapters to enable
`DEEPSEEK_API_KEY`	—	Required only when classifier=deepseek
`NLM_DISABLE_UPDATE_CHECK`	—	Set to `1` to disable the daily npm-registry update check
`NLM_RECALL_DECAY_HALF_LIFE_DAYS`	`180`	Half-life of the recency multiplier applied to recall scores. Older sessions score lower; defaults to 6 months. Set to `0` to disable recency weighting entirely.
`NLM_RECALL_DECAY_FLOOR`	`0.25`	Lower bound on the recency multiplier — even ancient sessions retain at least 25% of their raw score so a perfect-match old session can still surface.
`NLM_RECALL_REWRITE_DEFAULT`	`true`	Default value for the MCP `recall_sessions` `rewrite` parameter. When true, the service runs an LLM rewrite on vague natural-language queries before search. The HTTP hook caller bypasses rewrite regardless (hot-path protection).
`NLM_RECALL_REWRITE_TIMEOUT_MS`	`5000`	Per-call timeout for the rewrite LLM. Separate from the classifier timeout.
`NLM_FACT_CORROBORATION_BOOST_CAP`	`2.0`	Maximum multiplicative boost applied to fact recall scores based on how many sessions corroborate the same `(subject, predicate, value)`. Log-scale: 1 corroboration is 1.0×, 10 is 2.0× (capped). Set to `1.0` to disable the boost — the count is still returned on each hit.
`NLM_HOOK_INJECT_FACTS`	`true`	Whether to attach high-confidence facts about top-hit entities to the pointer block injected by the hook. Set to `0` to disable globally.
`NLM_HOOK_FACT_LIMIT`	`5`	Maximum number of facts in the "Known facts" section of the pointer block.
`NLM_HOOK_FACT_MIN_CORROBORATION`	`2`	Minimum number of sessions that must have asserted a fact before it qualifies for hook injection. Set to `1` to include single-source facts.
`NLM_HOOK_FACT_MIN_CONFIDENCE`	`0.7`	Minimum classifier confidence for a fact to qualify for hook injection.
`TELEGRAM_BOT_TOKEN` / `TELEGRAM_CHAT_ID`	—	Required for `nlm digest --telegram`

Config file

~/.nlm/.env — autoloaded by every CLI command. Mode 0600, owned by you, never readable by other users. The setup wizard writes the initial keys; you can edit it directly.

Reference producer

The pi quality-gate extension (in the pi-sandbox repo) is a ~10-line integration that emits nlm.signal per gate step and again on retry exhaustion. It is the canonical example of the session-embedded transport.

---

REST API

Daemon binds 127.0.0.1:3940 (override with NLM_PORT). Selected endpoints:

Method	Path	Auth	Purpose
GET	`/api/health`	Host-only	Liveness probe; returns `{version, status, service}`
GET	`/api/recall`	Bearer/Origin	Hybrid recall — `?q=`, `?mode=keyword\\|semantic\\|hybrid`, `?limit=`, `?include_superseded=true` (default off; opt-in to surface overturned sessions down-ranked, badged with `superseded_by`). Each result carries `status` + `superseded_by`.
GET	`/api/recall/stats`	Bearer/Origin	7-day stats: total, hit_rate, useful_hit_rate, top queries
GET	`/api/recall/recent`	Bearer/Origin	Last N recall events for live tail/telemetry
GET	`/api/recall/cite-stats`	Bearer/Origin	Citation rate over `?days=`
GET	`/api/session/:id`	Bearer/Origin	Full session body + supersedence links
GET	`/api/recall/facts`	Bearer/Origin	Structured fact search; includes `hint` field when empty
GET	`/api/facts/history`	Bearer/Origin	Version chain for one fact
GET	`/api/dataset`	Bearer/Origin	Full session list for the UI dataset view
GET	`/api/live/recent-writes`	Bearer/Origin	Live tail of ingested sessions
GET	`/api/data/backup`	Bearer/Origin	Streaming SQLite snapshot download
POST	`/api/data/restore`	Bearer/Origin	Stage a snapshot for apply-on-restart
POST	`/api/citation/explicit`	Bearer/Origin	Log an explicit session citation (the route backing the `cite_session` MCP tool; payload `{id, conversation_id?, reason?}`)
POST	`/api/hook/pre-compact`	Bearer/Origin	Hook endpoint; flushes the surfaced-IDs memo
ALL	`/mcp`	Bearer required	Streamable-HTTP MCP transport for container agents

/api/* is gated by three layers: 127.0.0.1 Host check (defeats DNS rebinding), Origin check when the browser sends one (defeats cross-origin drive-by), Bearer fallback when Origin is absent (server-to-server clients).

---

Pipeline

What happens when an AI runtime writes a session and you later recall it:

ingest:  runtime transcript (jsonl/sqlite)
   -> adapter parses runtime-specific format
   -> classifier (Ollama local by default; DeepSeek / OpenAI / Anthropic / OpenRouter / OpenAI-compatible if you opt in) extracts label + entities + decisions + open questions
   -> embedder (nomic-embed-text via Ollama) computes 768-dim vector
   -> SQLite canonical store + FTS5 keyword index + sqlite-vec ANN index

recall: prompt / query
   -> tokenize + match scoring (label x3, entity-exact x4, decision x2, summary x1, phrase-bonus +5)
   -> hybrid: BM25-style keyword + vector cosine, fused by score
   -> select-top-N gate (per-fire cap 3, per-conversation cap 10)
   -> pointer block prepended to model context (hooks) or returned as tool result (MCP)

---

vs Alternatives

	nlm-memory	mem0	Letta / MemGPT	Built-in (`CLAUDE.md`)
Unit of memory	Whole session + extracted markers	Atomic facts	Graph nodes + edges	Static file
Cross-runtime	9 adapters, one corpus	Per-app SDK integration	Per-app SDK integration	Per-runtime config
Editable timeline	Sessions can be superseded, retired, aborted	Append-only fact log	Graph edits	Manual file edits
R@5 baseline	97.2% on 14mo corpus	published varies	published varies	n/a
External deps	SQLite + Ollama (local)	Postgres or Qdrant	Postgres	none
Hosted offering	none — local only	yes	yes	n/a
Account required	none	yes (cloud tier)	yes	none
Telemetry	none	yes	yes	none
License	Apache 2.0	Apache 2.0	Apache 2.0	—

The defining property is the editable timeline. mem0 and Letta append; NLM lets you reach back and mark a session as superseded by a newer one, retire one as no-longer-relevant, or flag one as aborted-mid-flight. The next recall surfaces the corrected version, not the stale one. A claim from 6 months ago can be patched today.

---

🎯 aiskill88 AI 点评 A 级 2026-06-12

高质量的MCP工具，为AI运算提供了便捷的内存管理