Wyrm

MCP文档知识

提供方 未标注

文档知识 未标注 未标注 社区 / 第三方 Claude DesktopClaudeCursorVS Code

基于MCP的AI代理持久记忆:本地优先存储、从过往失败中负向学习、混合召回,以wyrm-mcp发布在npm上。

一句话选型结论

按资料字段整理

Wyrm 属于文档知识类 MCP 服务器,传输方式未标注,鉴权方式未标注,可直接接入 Claude Desktop、Claude、Cursor 等,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类文档知识 传输方式未标注 鉴权未标注 部署方式未标注
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

文档知识
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

记忆

安装与配置

未标注 未标注
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
文档检索问答

在你的文档库里按问题找答案。

知识库读写

把新资料写进知识库,随用随取。

内容整理

把零散资料归纳成结构化笔记。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 已适配 Claude Desktop、Claude、Cursor,换客户端不用重新研究配置。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在文档知识场景里让模型直接调用外部能力的人。
  • 已经在用 Claude Desktop、Claude 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
Wyrm当前 未标注 未标注 未标注 无 未标注
dify 未标注 stdio · 本地进程 本地部署 有 2026-04-11
open-webui 未标注 stdio · 本地进程 本地部署 有 2026-04-11
JeecgBoot 未标注 stdio · 本地进程 本地部署 有 2026-04-11
Cherry Studio 未标注 未标注 未标注 无 2026-05-17
mempalace 未标注 stdio · 本地进程 本地部署 有 2026-04-11

服务器档案

文档知识
字段来自项目仓库与公开索引,中文说明由本站标注
提供方未标注上游标注的维护方 未标注
分类文档知识本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式未标注远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型未标注上游给出的部署形态
最近核验未标注上游最后核验时间

详细介绍

9.8 KB
来自项目仓库与上游目录的原文,未做删改

Persistent memory for AI agents, over MCP.

Ground truths, recorded failures that block repeats, decision causality, and hybrid recall.

A structured SQLite memory on your machine. No cloud, no separate LLM.

Website ·
Quickstart ·
Discussions ·
Changelog


What it is

Most AI coding sessions start from zero. Wyrm gives the agent a memory that persists across them. It keeps your project's decisions, conventions, open work, and dead-ends in a structured database on your own machine, and hands them back to the model at the start of the next session. An agent connected to Wyrm recalls what was decided last week instead of re-deriving it, and can be stopped from repeating an approach that already failed.

It speaks the Model Context Protocol, so it drops into Claude, Cursor, Copilot, Windsurf, and Codex without glue code.

Quickstart```bash

npm install -g wyrm-mcp # install
wyrm-setup # wire it into your AI clients, then restart them


> **If `npm install -g` fails with EACCES**, npm's global prefix is a directory you cannot write (often `/usr`). Install under a prefix you own instead, and put its `bin` on your PATH: `npm install -g --prefix ~/.npm-global wyrm-mcp`, then `export PATH="$HOME/.npm-global/bin:$PATH"`. `wyrm update` later updates that same install.

> **On npm v12+**, npm denies dependency install scripts by default, which skips `better-sqlite3`'s native build — the install *succeeds* but `wyrm` then fails with *"Could not locate the bindings file"*. Install with the build allow-listed instead: `npm install -g wyrm-mcp --allow-scripts=wyrm-mcp,better-sqlite3,onnxruntime-node,protobufjs` (`wyrm update` already does this for you). See TROUBLESHOOTING.md.

Then, from inside your client, ask it to call `wyrm_capabilities` to confirm the connection. The everyday loop is four steps the agent runs on its own once the habit sets in:```
prime    →  load the project's truths, quests, and dead-ends at session start
recall   →  retrieve what you know before re-deriving it
check    →  before retrying an approach, ask if it already failed
capture  →  store durable facts, lessons, and tasks as you go

Measured, not asserted

Every number Wyrm publishes comes from a benchmark committed to the source, reproducible on your own data. The negative-learning firewall and the recall lift are the two that matter most, and both are covered below.

The Firewall Receipt

The receipt is the measured proof of one claim: when a failure is on record, the firewall blocks the repeat, stays silent on novel actions, and drops the block once the anchored source drifts, deterministically and with no model in the loop. On wyrm-mcp 9.1.2, measured 2026-09-12, the deterministic bench scores recall 100% (32/32) on same-action and reworded repeats with precision 100% (0 false blocks of 16), and in the blind agent A/B the repeated-failure rate falls from 38.9% (14/36) to 8.3% (3/36), a 78.6% relative reduction with a bootstrap 95% CI of 45.5% to 100% (n=36 per arm). The method, the per-scenario table and the two limits the numbers carry are in docs/firewall-receipt.md; the deterministic receipt is regenerated on every push in CI as the artifact firewall-receipt-node22.x (workflow).

The failure firewall in twenty seconds: a failure is on record, the agent proposes the same deploy, and the PreToolUse hook refuses it with the recorded reason before the command runs. A variant with an added flag gets an advisory, not a block, which is the honest edge of the deterministic match. Every line in the clip is real hook output.

Why Wyrm

It remembers what failed, not just what worked

Most memory tools store successes. Wyrm also records dead-ends and blocks the repeat. You record a failed approach with wyrm_failure_record, and a later wyrm_failure_check surfaces it before the agent walks back into it. Across a session that stops re-litigating solved problems; across a fleet of agents, one worker's dead-end warns the rest, once.

Recall that finds things by meaning, not just keywords

wyrm_recall runs keyword search (FTS5) and semantic search over a vector index, fuses them, and reranks. It is hybrid by default, no configuration required, no account, and no hosted call: a small local embedding model downloads automatically the first time you run wyrm-setup. It also weighs recency, temporal cues, and confirmed reuse, so the memory that fits the moment ranks first. On a real-set LoCoMo benchmark committed in the repo, the deterministic no-LLM floor is recall@5 52.4% / recall@10 59.9%, and the bundled local model lifts that to recall@5 60.0% / recall@10 72.0% (in line with the published local-hybrid reference of 60.3% / 72.2%, see BENCHMARKS.md). For maximum accuracy, wyrm upgrade moves to hosted NVIDIA NIM retrieval (below).

Local-first, and honest about egress

By default no memory data leaves your machine. The database is a single SQLite file at ~/.wyrm/wyrm.db. The only default outbound calls are a daily npm version poll and, on installs with a license key, a signed revocation-list fetch — both off with WYRM_NO_VERSION_CHECK=1 and WYRM_LICENSE_REFRESH=0 respectively; the free tier's one-time model download is a plain fetch that sends nothing. docs/EGRESS.md in the repository enumerates every destination, and if you find an outbound call that is not on that table, that is a bug worth reporting. When you do opt into a hosted embedding path, Wyrm reports exactly what left and where, in a determinism receipt and on its health endpoint. The privacy claim is one you can verify from the runtime, not just the docs.

Every write leaves a receipt

Every memory write returns a structured receipt that says what happened to it: stored, queued for review, merged, aliased, or dropped, and why. The receipts are ledgered too, so wyrm digest --writes reconstructs a day's writes offline and wyrm_stats shows the outcomes and the review-queue depth. You can audit what the agent actually committed to memory, not just trust that it did.

Built for one agent or a fleet

Every memory is attributed to the agent and run that produced it, so a swarm of agents can share one accountable memory bus, with failures kept private to your account by default. A live event stream keeps devices in sync.

Untrusted input stays out of the context brief

Every artifact is tagged with where it came from: you, an agent, an import, or an untrusted source. Content on the untrusted lane is categorically withheld from the context briefs the model reads, whatever it contains, and imported content is detector-gated and marked. The brief-surface quarantine was hardened against a full garak red-team of 622 real jailbreak payloads: zero escapes from the untrusted lane by construction, and the detector flagged 80.7% of the rest with zero false positives on benign prose. That red-team runs as a CI gate.

NVIDIA NIM retrieval (optional)

Local vectors are on by default. NIM is the further step up, for embeddings and reranking, when accuracy is worth a hosted call. On the LoCoMo retrieval benchmark committed in the repo (2 conversations, k=10, measured September 2026), the default NIM embedding model nvidia/nemotron-3-embed-1b reached recall@1 39.9% and recall@10 76.7%; adding the NIM reranker (nvidia/llama-nemotron-rerank-vl-1b-v2) raised recall@1 to 55.5% across the same 301 questions, at about one extra second per query. Figures published here earlier were measured on models NVIDIA retired on 25 August 2026, and have been withdrawn. It is an explicit opt-in, off by default, and the egress is disclosed on every call.

The guided path is wyrm upgrade: a free API key, masked key entry, a live call to validate the key before anything is written, and an optional one-time reindex of existing memories on the new tier. To configure it by hand instead:bash export WYRM_VECTOR_PROVIDER=nim export WYRM_RERANK_PROVIDER=nim export NVIDIA_API_KEY=nvapi-...

Ghost Protocol (Pvt) Ltd is a member of NVIDIA Inception.

Works with

Claude (Code, desktop, web) · Cursor · GitHub Copilot · Windsurf · Codex, and any MCP-capable client.

Requirements

Node.js 22 or newer. Semantic recall is on by default via a small bundled local model (wyrm-setup downloads it, no account needed); wyrm upgrade adds hosted NVIDIA NIM retrieval for maximum accuracy. Already running Ollama for this? Existing setups keep working unchanged.

Pricing

Wyrm is free to use forever — the local memory, the firewall, the recall, all of it. Paid plans add cloud sync across devices, encrypted snapshots, shared team memory, and support:

  • Pro — $29/mo · cloud sync, encrypted snapshots, analytics, priority support
  • Team — $199/mo · shared team memory, up to 25 seats, admin dashboard
  • Enterprise — $499/mo · unlimited seats, SSO/SAML, custom SLA, on-premise option

Plans and self-serve checkout: account.ghosts.lk/pricing. Custom or on-premise, email ryan@ghosts.lk.

Community and feedback

Wyrm sends no telemetry, so the way we learn what is working is you telling us.

  • wyrm feedback from your terminal opens a prefilled report. Add --bug, --idea, or --question.
  • Discussions for questions, ideas, and how it is going.
  • Issues for bugs and concrete requests.
  • Email ryan@ghosts.lk.

If Wyrm earned a place in your workflow, a star helps other people find it.

License

Wyrm is proprietary software, free to use under the Wyrm license and Terms of Service. This repository is the public home for the docs, the changelog, and the community. The source is not published here. For a commercial license (embedding Wyrm in a closed product, or running it as a managed service), contact ryan@ghosts.lk.

Built by Ghost Protocol · Colombo, Sri Lanka

NVIDIA, the NVIDIA logo, and NVIDIA Inception are trademarks and/or registered trademarks of NVIDIA Corporation.

常见问题

6 问
答案由本站依据条目字段整理,不引用上游原文
Wyrm 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

Wyrm 怎么在 Claude Desktop、Claude、Cursor 里配置?

它的传输方式未标注,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

Wyrm 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

Wyrm 需要 API Key 吗?

资料没有标注鉴权方式,是否需要密钥取决于服务本身,接入前按官方文档确认。

Wyrm 支持哪些客户端?

资料标注支持 Claude Desktop、Claude、Cursor、VS Code、Windsurf、Gemini CLI;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

Wyrm 能提供哪些能力?

资料显示能力标签包括 记忆。具体工具清单以官方文档为准。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 未标注。