官方 ContextStream

MCP文档知识

提供方 未标注

文档知识 未标注 未标注 官方维护 Claude DesktopClaudeCursorVS Code

跨会话的AI编程助手的持久化记忆与语义搜索

一句话选型结论

按资料字段整理

官方 ContextStream 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 官方维护 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
官方 ContextStream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
维护方官方维护核验等级 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

19.3 KB
来自项目仓库与上游目录的原文,未做删改

Agents Remember

Git-verified records for what your coding agents know. A control plane for what they do.

📖 Current docs: https://foxfire1st.github.io/agents-remember/
🤖 Machine-readable summary: https://foxfire1st.github.io/agents-remember/llms.txt
Note: caches and search snippets may serve an outdated copy of this README — the docs site above is canonical and always current.

Table of Contents

  1. Why It Exists
  2. Core Features
  3. What It Looks Like In Practice
  4. Live Demo
  5. Requirements
  6. Quickstart
  7. Run The Dashboard
  8. Documentation
  9. Repository Layout
  10. Status
  11. Stability
  12. Contributing

Why It Exists

Modern coding agents can make clean, plausible edits while missing the project-specific rules that make those edits safe. A top-level instruction file can help, but it does not naturally reappear when the agent is deep in a file and deciding what to change.

Agents Remember fixes that: the matching note is reachable at the moment of the edit — most often by the very path the agent is already working in — so project rules surface exactly when a change is being made, not buried in a top-level file.

Core Features

Agents Remember gives coding agents project memory they can verify and act on. It turns local invariants, naming rules, migration scars, cross-repo contracts, and "this looks safe but is not" facts into versioned Markdown beside the code, checks that memory against Git before use, and updates it only after approved work lands.```text
src/orchestrator/core_editor.py
ar-memory/onboarding/src/orchestrator/core_editor.py.md


- **Path-addressed memory:** A source file's note lives at a deterministic mirror path, so an agent holding a file can reach the right context without search, ranking, or guesswork.
- **Git-proven freshness:** File notes, route overviews, and entity catalogs are drift-checked against source commits, route scopes, or deterministic fingerprints before they are trusted.
- **Search that finds, not decides:** Optional semantic memory and code-graph providers help locate relevant files, callers, dependencies, and concepts, but verified Markdown and source code remain the truth.
- **Memory that lands with code:** External memory repos use a `memory.md` ledger, isolated dual worktrees, preview/apply closeout, and all-or-nothing integration so code and memory stay synchronized.
- **Repo-owned agent behavior:** Each memory repo carries `system/` files for path rules, tools, coding guidelines, documentation sources, branch policy, and reporting shape, so the same project rules load across harnesses.
- **Harness-ready first run:** Starter packages for Claude Code, Codex, Cursor, Antigravity, VS Code Copilot, Hermes, Pi.dev, and OpenClaw carry the native MCP, skills, hooks, rules, and instruction files each harness needs.

The default setup stores durable memory in the target repository under `ar-memory/`. Teams that need separate memory repositories can use external memory under `ar-coordination/memory-repos/ar-/`. For the full tour, see Features.

## What It Looks Like In Practice

A source file has an onboarding note beside it, reached by path:```text
mcp/src/agents_remember/mcp/server.py
ar-memory/onboarding/mcp/src/agents_remember/mcp/server.py.md

At task start the agent orients and checks memory health:```text
context_packet(repo_id="my-app")
memory_quality_check(request={"mode":"sync", "repo_id":"my-app"})


It then reads the source file and its onboarding note together before proposing a change. After the change is approved and lands, the onboarding is refreshed and re-verified against the new commit — so the note stays true to the code.

## Live Demo

Agents Remember runs on itself. The companion memory repo is:
https://github.com/Foxfire1st/ar-agents-remember

That repo contains the live onboarding layer, so you can inspect how by-path memory, drift-aware updates, and contribution-time onboarding look in practice.

## Requirements

Before the Quickstart, make sure the host has:

- **uv** (for `uvx`) or pip, and **Python 3.13** — the package supports `>=3.13,<3.14`; repository development uses the verified source-built 3.13.15 contract documented in the MCP README.
- **Git**, with `user.name` / `user.email` configured (memory and worktree commits need an author; otherwise a placeholder identity is used).
- **Docker** running, only if you enable the optional providers. The semantic-memory provider (grepai) also uses a Dockerized Ollama and pulls an embedding model (`nomic-embed-text`) on first setup — no host Ollama install needed.

Providers, Docker, and Ollama are only needed for the optional Docker-backed
providers; the core by-path memory works without them. Claude Code hooks do not
require `jq`; the current starter package uses a Python hook. Full detail and
troubleshooting live in the MCP package README.

## Quickstart

This is the short path for a new workspace. The detailed walkthrough lives in Getting Started.

Ask your agent to:

1. **Copy the harness package** — Pick your harness guide under
   docs/install, copy that harness's native starter
   files from this repo into the workspace, then render the copied package.
   The `render-starter` script is a convenience: it infers the workspace root
   from the copied harness folder and fills the copied package's path,
   repository, and hook-command placeholders from a single `--repo` list such as
   `--repo my-app shared-lib`. You can also do those replacements by hand. These
   packages include the harness-visible skills, hooks/rules/instructions, and
   MCP settings templates.
2. **Wire the MCP server** — Register Agents Remember MCP from
   PyPI with `uvx`:```text
   uvx agents-remember-mcp@latest --config /absolute/path/to/agents-remember-settings.json
   ```

   Use the `agents-remember-settings.json` path from the copied harness package.
   Then **restart the harness once** so it loads the MCP server, native skills,
   and package hooks/rules/instructions.
3. **Onboard your project** — Invoke the copied skill
   `c-13-install-and-onboard`. It runs or verifies `runtime_install()`, asks
   whether to scaffold a new memory repo or use an existing one, bootstraps
   onboarding when needed, and starts provider indexing when providers are
   enabled.

That is the normal first-run path. `skills_install()` remains available as a
maintenance/manual MCP tool, but the starter packages already provide the
initial skills and harness files.

After that, normal work runs through the `l-01-agent-lifecycles` skill: developer-facing free chat answers research inline and, for ordinary role-shaped work after the durable sprint and first leaf exist, compiles the canonical architect brief and calls `dispatch_agent` once on that sprint document. An explicit developer-declared task-seat takeover instead targets the named role on its canonical task document. The identity-free launcher hands over after the exact brief is durable; later plane-hosted seats use the same tool under structural child-scope authority. Spawned backend seats follow their role briefs. The agent resolves the active context with `c-08-ar-coordination-context-resolver`, checks memory quality with `c-02-memory-quality-control`, reads relevant onboarding beside code, and updates onboarding after approved changes.

## Run The Dashboard

The mission-control dashboard ships inside the MCP package. Install the CLI
once with uv — latest stable, no version pin — then start the cockpit from
anywhere in your workspace:```text
uv tool install agents-remember-mcp
agents-remember dashboard

--config is optional: the CLI walks up from the current directory and uses
the nearest .claude/mcp/agents-remember-settings.json, or the --config
recorded in an .mcp.json agents-remember entry — the same settings file
the MCP server boots from.

For a dashboard that survives closing the terminal, use daemon mode:```text
agents-remember dashboard --daemon # detach; state + log under <coordinationRoot>/logs/dashboard/
agents-remember dashboard --status # exit 0 when running, 1 when not
agents-remember dashboard --stop


Or let the MCP server supervise it: set `&quot;dashboard&quot;: {&quot;autoStart&quot;: true}` in
the MCP settings JSON and every server boot ensures the daemon — adopting a
healthy one, starting a missing one, and restarting on version mismatch so an
upgrade is picked up by the next session
(Settings Reference).

Pinning a version is the debugging/repro path, not the default: `uv tool
install &#x27;agents-remember-mcp==3.0.0rc8&#x27;`, or one-shot without installing,
`uvx --from &#x27;agents-remember-mcp==3.0.0rc8&#x27; agents-remember dashboard`.

&gt; **Pre-release note (until 3.0.0 final):** the dashboard currently ships in
&gt; `3.0.0rcN` pre-releases, which default version resolution skips. Install with
&gt; `uv tool install --prerelease allow agents-remember-mcp`, and register the
&gt; MCP server with an explicit `agents-remember-mcp==3.0.0rcN` pin instead of
&gt; `@latest`.

## Documentation

- Features - the concentrated tour of what Agents Remember gives users.
- Getting Started - a fuller first-run setup.
- Concepts - onboarding units, memory roots, drift, and approval gates.
- Architecture - runtime, coordination, internal memory, and external memory.
- Workflows - the `l-01-agent-lifecycles` skill and its build modes (research-only exit / `w-02-light-task-workflow` skill task / master + light sub-task series), and when to use each.
- Benchmark Methodology - how paired `codex exec --json` runs are captured and compared.
- FAQ - design principles, objections, and comparisons.
- External Memory Guide - separate memory repos for selected code repos.
- Cost-aware Bootstrap - model and wave-sizing choices for token-heavy repository bootstrap.
- Settings Reference - memory-layer `system/settings.json` and MCP authority settings.
- Skills Reference - the installed skill families.

## Repository Layout```text
agents-remember/
  AGENTS.md                         # source checkout instructions
  README.md                         # public front door
  skills/                           # canonical skill source tree
  scripts/sync-skills.py            # sync skills into package/harness copies
  scripts/sync-runtime.py           # sync runtime assets into package data
  scripts/sync-harness.py           # generate the nine harness configuration trees
  scripts/harness/                  # canonical source for those trees
  agents-md-files/                  # canonical installed AGENTS.md templates
  benchmarks/                       # canonical optional benchmark package source
  providers/                        # canonical provider runtime assets
  system/defaults/examples/         # canonical scaffold examples
  mcp/                              # package-local MCP server and services
    src/agents_remember/package_data/
      runtime/
        agents-md-files/            # generated copy of root agents-md-files/
        skills/                     # generated package copy of root skills/
        providers/                  # generated copy of root providers/
        system/defaults/examples/   # generated copy of root system/defaults/examples/
      benchmarks/                   # generated copy of root benchmarks/
  docs/                             # user-facing documentation

Edit skills in root skills/, then run python3 scripts/sync-skills.py to
refresh the MCP package data and every harness starter package. The pre-commit
and pre-push hooks run python3 scripts/sync-skills.py --check.

Edit runtime assets in root agents-md-files/, benchmarks/, providers/,
and system/, then run python3 scripts/sync-runtime.py to refresh MCP package
data only. The pre-commit and pre-push hooks run
python3 scripts/sync-runtime.py --check.

Edit the self-hosted harness configuration in root scripts/harness/, then run
python3 scripts/sync-harness.py to regenerate the nine .claude/, .codex/,
.cursor/, .github-vscode/, .vscode/, .hermes/, .openclaw/, .pi/ and
.agents/ trees. The pre-commit and pre-push hooks run
python3 scripts/sync-harness.py --check, and mcp/tests/test_sync_harness.py
runs the same check inside the suite.

The hooks are tiered, and both are thin wrappers over .githooks/_gate.sh.
pre-commit runs the fast tier over the staged content: the generated-copy
checks above, plus Ruff, ruff format --check, Pyright, and deterministic dashboard
checks. pre-push repeats those non-test checks against current-checkout bytes and
records the pushed refs. It does not run acceptance. GitHub runs its deterministic
non-test checks once per pull request, not once again for every branch push. A
pinned Dagger v0.21.8 graph rebuilds the exact Git candidate
in a clean Ubuntu container, installs from scratch, runs a bundled real Codex
read-only protocol probe, and executes the accepting wrapper. Targeted Dagger runs
once when each leaf closeout creates its commit. Leaf integration lands that exact
certified commit without a rerun. Full Dagger runs once when each master integrates
into super. PR validation, tagging, and publishing do not rerun acceptance. See
CONTRIBUTING.md for the tier table and staged-content contract.

Agents Remember declares that Dagger graph in its repository-owned
mcp/certification-profile-v1.json, selected explicitly by
repositories.agents-remember.certificationProfile in the MCP authority settings. The framework
does not discover a wrapper or carry an Agents Remember command/report inventory. Ordinary Python development uses pytest directly, without Dagger admission, coverage,
repository certification, or an autouse application service graph:```bash
mcp/.venv/bin/python -m pytest # default unit loop
mcp/.venv/bin/python -m pytest mcp/tests/test_example.py # one changed behavior
mcp/.venv/bin/python -m pytest -m integration # delivery boundary checks


The default excludes the `integration` marker. Local inputs, temporary resources, and
explicit test doubles remain ordinary tests; real publication/recovery, competing writers,
application wiring, and whole-repository observations run separately. Imported test classes
are exercised only in their defining module. Four workers are the default; use `-n=0` for
serial debugging. Tests use disposable home/config/cache directories and scrub inherited
Git selectors, live opt-ins and credentials. They never declare a daemon identity.

Delivery runs both populations together (`-m &quot;&quot;`) in the existing shared Dagger environment.
Only explicit `--certify` loads the retained certification plugins and requires genuine
Dagger admission. Combined branch coverage feeds the **90% changed-production-code floor**
and the existing **CRAP threshold of 30**. Tests and verification-only support are excluded
from production scoring. Coverage and CRAP are not part of the ordinary unit command.
Direct targeted Vitest unit/component commands also remain available.

The full evidence taxonomy, lifecycle metadata, fixture-authority rule, dependency-owned
selection/retry behavior, stress cadence, and causal-failure contract are documented in
`docs/design/python-evidence-system.md`.
Leaf/focused acceptance is Dagger `mode=targeted`, while the single master-altitude
full-repository acceptance is Dagger `mode=full`. Both require an explicit Git
`diff-base`; the public Dagger function refuses an empty base instead of comparing
the candidate to Git&#x27;s empty tree. Run `dagger call quality --help` to see the
current modes and argument contract. There is no direct-Docker or host fallback:
an unavailable Dagger engine fails explicitly.
The graph receives a separate Git ancestry bundle plus the exact staged source,
never the live coordination root, credentials, or container socket. Its live
trace and final pytest, coverage, Codex-probe, and result artifacts replace the
corresponding files under the task enclosure&#x27;s `reports/` directory.

Inside the nonce-attested Dagger graph, the wrapper orders cheap deterministic rails before the
expensive test rail: Ruff, formatting, file size, Pyright, Radon reports, then pytest. CRAP and
changed-lines coverage score that run&#x27;s branch-coverage artifact last. Content-addressed exact or
test-only retry proof is an internal Dagger optimization; any source, configuration, selected-suite,
runtime, environment, or artifact drift runs the ordinary selection in the same graph. There is no
host retry path or fallback. Lifecycle acceptance disables proof reuse by default with
`AR_QUALITY_NO_RETRY=1`.

Every Dagger rail prints one provenance line naming its actual input, resolved config, and unit
count. The deterministic pre-push hook separately forwards Git&#x27;s ref updates and checks
current-checkout bytes at index-known paths; it runs no tests and never claims acceptance for the
pushed commit range.

The installed runtime lives in `ar-coordination/` — by default `/ar-coordination/`,
inside the workspace (never your home directory) — not in the source checkout. The
`c-13-install-and-onboard` skill shows this and every other install path as a workspace-first
default you can accept or override:```text
ar-coordination/
  AGENTS.md
  skills/
  system/
  memory-repos/
  providers/                        # provider runtimes (images, runners, indexes)
  benchmarks/                       # optional, installed with --include-benchmarks
  tasks/
  notes/
  worktrees/
  temp/

Status

Agents Remember is at 3.0.0rc8 and actively developed. The core path — by-path onboarding, drift checks, and approval-gated updates — is in real use and stable enough to rely on. The public contracts listed under Stability are held stable across minor releases and change only on a major bump; the internals beneath them and the optional semantic/relationship providers may still evolve, so pin a version and read the notes for your target version in GitHub Releases — the repository's canonical changelog — before upgrading. The Claude Code path is the most exercised; other harnesses are supported but less battle-tested.

The 3.0 arc: the working session itself is now observable and steerable — a system-managed agent lifecycle with durable approval gates and an event/projection layer, served as the mission-control browser cockpit directly from the MCP package (agents-remember dashboard; #2, #43). The rc tag means the cockpit surface is still settling toward the final 3.0.0 contract; the architecture beneath it is the one described above.

Stability

Following semantic versioning from 1.0.0, these public contracts will not change without a major version bump: skill IDs (e.g. the c-08-ar-coordination-context-resolver and w-02-light-task-workflow skills), MCP tool names and their inputs/outputs, the ar-coordination/ and ar-memory/ layout, and the settings schema. Internal modules, provider internals, and prompt wording are not part of this promise and may change in minor releases.

Contributing

Contributions should make the memory layer clearer, safer, and easier to apply consistently. Start with CONTRIBUTING.md and keep the core rules intact: drift check before planning, approval before implementation, and onboarding updates only after approved changes.

Agents Remember runs on itself, so the best way to contribute is with the memory layer active. Download or clone this project's own memory at Foxfire1st/ar-agents-remember and use it as the Agents Remember memory for your checkout: you get the project's by-path onboarding at the moment you edit, and your onboarding updates land alongside your code changes — the same loop this repo asks of every contribution.

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
官方 ContextStream 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。

官方 ContextStream 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。

官方 ContextStream 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。

官方 ContextStream 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。

官方 ContextStream 支持哪些客户端?

资料标注支持 Claude Desktop、Claude、Cursor、VS Code、Windsurf、Cline;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。

官方 ContextStream 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。