AI Memory Book

MCP文档知识

提供方 未标注

文档知识 未标注 会话登录 社区 / 第三方 ClaudeCursorVS CodeWindsurf

面向编程代理的本地 MCP 记忆:基于项目范围的 SQLite/FTS5 检索原始对话证据,并提供来源页面引用和有界上下文。

一句话选型结论

按资料字段整理

AI Memory Book 属于文档知识类 MCP 服务器,传输方式未标注,会话登录,可直接接入 Claude、Cursor、VS Code 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 鉴权会话登录 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 会话登录
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、Cursor、VS Code,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。
  • 使用前要准备好会话登录对应的凭据,并按最小权限配置。
  • 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、Cursor 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

  • 不方便管理密钥或第三方授权的人。
  • 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
AI Memory Book当前 未标注 未标注 未标注 无 2026-11-20
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
鉴权方式会话登录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2026-11-20上游最后核验时间

详细介绍

23.3 KB
来自项目仓库与上游目录的原文,未做删改

AI-Memory-Book

Give AI an addressable book of engineering memory, with a separate book for each project.

Preserve conversations as project books, task chapters, question sections and original evidence pages. Find a chapter from its title or question, then follow its addresses to the source. Similar codebases can reuse task names and parameters while keeping their records in explicit project namespaces. Recover decisions, corrections, constraints and failed attempts when a session ends or a client changes.

v0.13.0 · MIT · Python 3.10+ · SQLite / FTS5 · Optional native C acceleration

v0.13.0 is published. The independently audited long-history replay reaches 100,000 pages across 788 projects. At every one of seven checkpoints, the v0.13 source returns complete evidence for 56/56 answerable recall cases, compared with 48/56 for the v0.12 baseline. Directory navigation has lower pooled p50 in all 14 measured query-stage groups; recall has both improvements and slower cases. The release also passed its own final-source CI, native build/verification matrix and public-URL installation. Results and limits · v0.13 publication proof.

At a glance

Complete evidence Directory contracts Current-declaration diagnosis Targeted safeguards
72 / 72 answerable cases 119 / 119 checks per backend 38 / 38 answerable cases 37 / 37 controls per backend
Original 80-question engineering fixture; two budgets and two packing tracks Navigation, pagination, exact scope and original-page reads Separate 62-question diagnostic; four auto-mode budget/limit settings Quotation, revocation, field/scope identity and independent report facts

These are authored development-fixture results with distinct denominators. They measure retrieved evidence and API contracts, not coding-model answer accuracy, independent holdout quality or production task success. Exact measured sources, raw records and limits.

What you need What the source implementation provides
Resume an engineering task Original decisions, later corrections, constraints and failed attempts with source references
Keep similar projects separate Exact project IDs and project-bound MCP connections, including follow-up page reads
Locate a task from its heading Indexed title/question/directory search, unique chapter results and scoped follow-up addresses
Recover a later decision Check complete source-local declarations, field identity and environment before current-state ordering; retain historical/event evidence when requested
Recognize uncertainty Surface bounded comparable-declaration and potential-conflict diagnostics, with original page IDs for inspection
Search code and conversations Chinese/English lexical retrieval plus acronym, camel-case and snake-case aliases
Control the returned context A default 2,400-character budget and up to five results; optional exact token caps
Switch clients One archive shared across supported Claude Code, Codex, Gemini and other integrations
Operate locally SQLite / FTS5; default ingestion and lexical recall make no model calls
Recover safely Verified online backups, transactional source replacement and snapshot-based export

Quick start

Install the published v0.13.0 release with the pinned commands below. Its release run and real public-URL installation completed successfully.

Requires Python 3.10+ with SQLite FTS5. The installer selects a compatible prebuilt native wheel or the portable Python fallback and creates an isolated environment with MCP/client dependencies.

macOS / Linux```bash
curl -fsSLo install-memory-book.sh https://raw.githubusercontent.com/Afloat16/AI-Memory-Book/v0.13.0/scripts/install.sh
bash install-memory-book.sh


**Windows PowerShell**```powershell
Invoke-WebRequest https://raw.githubusercontent.com/Afloat16/AI-Memory-Book/v0.13.0/scripts/install.py -OutFile install-memory-book.py
py -3 -I -X utf8 .\install-memory-book.py

Connect and check```bash
memory-book setup --client claude-code
memory-book --version
memory-book doctor


Choose your supported client profile. Use the full launcher path printed by the installer if it is not on PATH, then restart the client.

| Optional setup | Meaning |
|---|---|
| `--mcp-project YOUR_PROJECT_ID` | Bind that MCP connection to an exact stored project ID |
| `--capture` | Opt in to existing and future supported text transcripts |
| `--capture --background` | Also install and start background capture for a supported profile |

MCP configuration and native capture have separate support matrices. Tool invocation remains controlled by the host/model. Existing schema-4 archives need no new migration; preserve custom installation paths, refresh setup and restart existing processes when upgrading. Upgrade guide.

## Evidence coverage

### Long histories across similar projects

The same frozen protocol compares v0.12 and v0.13 across seven stages: a 1,024-page reference, a timestamp-only age shift, foreign-project growth to 10,016 and 50,016 pages, target-history growth to 100,000 pages, interleaved insertion of the same logical sources, and source replacement. The run completed **13,608 API measurements**; a separate offline audit verified the saved responses and 66,136 source references. These are authored development cases, including cases that informed the fixes. Original artifact and restoration · Independent audit.

| Quality outcome across seven stages | v0.12 baseline | v0.13 source |
|---|---:|---:|
| Complete answerable recall configurations | 336 / 392 | **392 / 392** |
| Required source spans recovered | 392 / 448 | **448 / 448** |
| Current source first, where evaluated | 80 / 168 | **112 / 168** |
| Stale source ahead of current source | 32 | **0** |
| Complete directory evidence | 56 / 56 | **56 / 56** |
| Missing-evidence queries still returning context | 56 / 56 | **56 / 56** |
| Source errors / cross-project output | 0 | **0** |

The same 56 answerable queries repeat across stages; 392 is not an independent-question count. Every stage now includes both required sources for all eight multi-source questions, with the boundary constraint first and current policy second. That preserves the original scoring: **complete evidence does not mean the current source always ranks first**. Missing-evidence queries still receive related context, so reliable abstention remains unproven. The original failed run and all intermediate regressions remain recorded. Full interpretation and costs.

### Original fixed engineering workload

**Fixed workload:** 8 projects · 4 algorithm families · 6 chronological checkpoints · 80 questions, including 72 answerable cases and 8 missing-evidence queries. The v0.12 baseline and final v0.13 source each completed **800 scored response configurations**: 320 Book configurations and 480 repository-strategy controls. Independent source, budget and score checks found **zero score changes** between these two versions. At both 1,200 and 2,400 characters, Book retains **80/80 required spans, 72/72 complete cases and 48/48 fixed-policy checks** in both its native-excerpt and shared whole-block tracks. These are source results on an authored development fixture. Exact source identities, every score and limits.

### Compared with practical retrieval strategies

At 2,400 characters, using each strategy's native context assembly:

| Strategy | Required spans | Complete evidence | Fixed-policy checks | Mean context characters |
|---|---:|---:|---:|---:|
| **AI-Memory-Book** | **80 / 80** | **72 / 72** | **48 / 48** | **1,798.9** |
| Scoped flat FTS | 60 / 80 | 52 / 72 | 28 / 48 | 1,801.3 |
| Recent history within budget | 64 / 80 | 56 / 72 | 32 / 48 | 2,260.9 |
| No memory | 0 / 80 | 0 / 72 | 0 / 48 | 0 |

**20 more complete-evidence cases than flat FTS; 16 more than recent history, using 20.4% fewer context characters.** Recent history uses shorter citation headings and has no top-five candidate limit. These are repository strategies, not additional external products. Protocol and current results.

### Compared with pinned open-source components

Same 80 questions, six chronological checkpoints, original supplied-page hashes, exact project scope, at most five candidates and a shared **2,400-character whole-block packer**. Book is replayed on final v0.13 source `91412b43`. The external rows retain their original pinned executions, with matching workload and supplied-page digests. **None of these external components was rerun in the v0.13 100,000-page experiment.**

| Component / tested mode | Required spans | Complete evidence | Fixed-policy checks | Stale-only cases | Mean packed characters |
|---|---:|---:|---:|---:|---:|
| **AI-Memory-Book** | **80 / 80** | **72 / 72 · 100%** | **48 / 48** | **0** | **1,798.9** |
| BM25S 0.3.16 · retrieval library, native defaults | 80 / 80 | 72 / 72 · 100% | 48 / 48 | 0 | 1,846.7 |
| Whoosh 2.7.4 · BM25F, configured OR parser | 80 / 80 | 72 / 72 · 100% | 48 / 48 | 0 | 1,457.8 |
| Haystack 3.5.0 · InMemoryBM25Retriever | 74 / 80 | 66 / 72 · 91.67% | 42 / 48 | 0 | 1,884.7 |
| LanceDB 0.41.0 · native FTS | 72 / 80 | 64 / 72 · 88.89% | 40 / 48 | 4 | 1,876.7 |
| billion-context · pinned search kernel | 72 / 80 | 64 / 72 · 88.89% | 40 / 48 | 0 | 1,955.0 |
| claude-mem · raw-observation FTS | 72 / 80 | 64 / 72 · 88.89% | 40 / 48 | 0 | 1,488.5 |
| Basic Memory · explicit FTS | 72 / 80 | 64 / 72 · 88.89% | 40 / 48 | 0 | 1,333.1 |
| Mem0 / Chroma / LangGraph · shared MiniLM configurations | 55 / 80 | 47 / 72 · 65.28% | 25 / 48 | 4 | 1,963.1 |

**Book, BM25S and configured Whoosh tie on complete evidence and fixed-policy checks at both budgets. Whoosh uses fewer mean packed characters than Book at both budgets.** At 2,400 characters Book finds six more complete cases than Haystack, eight more than LanceDB and each of the three historical lexical modes, and 25 more than each shared-MiniLM configuration. This does not establish universal superiority.

Mem0 uses `infer=False`; claude-mem's observer and Basic Memory's full default hybrid workflow are not run. The three MiniLM interfaces execute separately but share a representation and have correlated scores. Whoosh uses an explicitly configured **OR** parser, not its default AND parser. Haystack, BM25S, LanceDB and Whoosh are framework/store/retrieval components. These are configured component comparisons, not full-product rankings. Modes, pins and source audit · LanceDB/Whoosh execution and original records.

**Memvid SDK 2.0.160, lexical mode, was also executed:** every original natural-language question returned an empty native hit list. A separate 14-call diagnosis recovered indexed pages for original-text keyword controls, while the full question still returned no hits with or without project scope. The formal responses are retained as a configured-query limitation, not a zero-score product victory or an abstention result. No rewritten query replaces a scored question. Complete extension, audit and diagnosis.

The architecture review covers **26 external project families**, including six store/retrieval libraries or frameworks. **Eleven external configurations have been executed:** ten are represented in the scored table, while Memvid's query limitation is separate. The remaining families, including Hindsight, MemSearch, MCP Memory Service, Letta, Graphiti, MemOS, Memori, Cognee, OpenViking, Serena, EverOS, SimpleMem, Supermemory, LongMemory and the official MCP memory server, receive official-source feature coverage without invented numerical results. Architecture, deployment and upstream references.

### At the 1,200-character budget

| Book v0.13 source / packing | Required spans | Complete evidence | Fixed-policy checks | Stale-only cases | Mean characters |
|---|---:|---:|---:|---:|---:|
| Native excerpts | 80 / 80 | 72 / 72 | 48 / 48 | 0 | 1,196.6 |
| Shared whole-block packer | 80 / 80 | 72 / 72 | 48 / 48 | 0 | 1,041.5 |

These scores are unchanged from the v0.12 baseline. At the same 1,200-character common-block boundary, **BM25S and configured Whoosh also have 80/80 spans, 72/72 complete cases, 48/48 policy checks and zero stale-only cases**; their original mean lengths are 1,045.3 and 884.5 characters. The earlier v0.11 → v0.12 improvement remains in the historical v0.12 results.

The two packing tracks remain separate. Zero stale-only means that no case contains a registered stale span while lacking every required span; old and current text can coexist. Book and all ten scored external configurations return related context for the eight missing-evidence queries. Repeated maintenance text, no independent holdout and no coding LLM limit the conclusions. These scores do not establish calibrated abstention or general factual authority. All scores and boundaries.

## Performance

The same long-history replay measures seven recall queries and two complete directory chains per stage, using six balanced rounds and a separately restored baseline-built database for each arm. Each recall query-stage/version has 30 cold-result-cache calls and 90 warm-result-cache calls; directory chains have 30 calls. “Cold” clears the recall-result cache, **not** SQLite or operating-system caches. The single hosted Linux runner uses Python 3.12, SQLite 3.45.1 and an AMD EPYC 7763 CPU. Protocol, full timings and resource observations.

| Measured query-stage groups | Lower v0.13 pooled p50 | Higher v0.13 pooled p50 |
|---|---:|---:|
| Cold-result-cache recall | 12 / 49 | **37 / 49** |
| Warm-result-cache recall | 42 / 49 | **7 / 49** |
| Complete directory navigation and page reads | **14 / 14** | 0 / 14 |

At the 100,000-page interleaved checkpoint, archived-project current recall changes from **166.898 → 102.478 ms**, historical recall from **133.222 → 67.912 ms**, and the archived-project directory chain from **159.154 → 26.707 ms** (pooled p50). These measured gains coexist with costs: the worst cold p50 increase is **56.4%** for a 1,024-page multi-source query; a small archived-project current query is also **56.3% slower**. All seven warm slowdowns are multi-source queries, whose previously missing evidence is now included; their extra latency remains reported.

The figures show pooled p50 and nearest-rank p95. The complete 112-group comparison and resource record also retain six-round paired ratios, initialization, storage and process-lifetime peak RSS. Pooled and paired direction counts can differ; every slower group remains listed.

The directory and chronology routes reuse matching work while preserving ownership checks, source IDs and result order. Small project membership can avoid an unhelpful row-ID span, with fallbacks for other scopes and SQLite constraints. **Cold recall still depends on archive size.** Bounded returned candidates and exact project filters do not bound every database scan or global BM25 statistic. Snapshot restoration, opening the database, scoring and journaling are outside the query timer; these are API measurements, not coding-LLM or end-to-end task latency.

The optional C backend retains a complete Python reference fallback. Earlier v0.10 semantic/directory experiments and native-core experiments keep their original timings, costs and slower cases. Those historical speedups are not assigned to v0.13 or added to this experiment.

## How it works

| Step | Behavior | Agent interface |
|---|---|---|
| Archive | Keep original turns and coordinates; group by project/task/session | Import, capture, `remember_session` when enabled |
| Locate | Search indexed headings and questions, group by chapter, and retain project scope | `search_directory` |
| Find evidence | Apply project scope before candidate truncation; search original pages and identifier aliases | `recall_memory` |
| Inspect | Browse only when helpful; follow citations to original pages | `browse_book`, `browse_chapter`, `browse_section`, `read_page` |
| Continue | Verify constraints, corrections and failed attempts before editing | Your existing coding client |

The default budget includes citation headings and separators inside `context`; it does not cap the complete JSON response or model tokens. `temporal=auto/current/relevance` controls evidence ordering; source inspection remains necessary when judging authority.

### Find the chapter, then inspect the evidence```bash
# Use an exact stored project ID and a phrase from a task or question heading.
memory-book locate "gradient accumulation" --project service-alpha --limit 10

The locator searches metadata even when the original body does not repeat the heading. It returns unique chapters, each with at most five matching section previews, and explicit result_type="directory", evidence=false. Follow returned IDs with browse_chapter, browse_section or read_page; titles and summaries are navigation hints, not answer citations. Any analyzed metadata term can match, so a partial heading match may be returned.

The authored eight-project directory workload was rerun on each of the Python and native backends. Each passed 119 contracts, with 130 locator calls and 161 original-page reads covering 80 distinct pages. The original protocol correction and all attempts remain recorded. This is a navigation/scope result, not answer accuracy or a speed benchmark. Final v0.13 replay and independent source audit.

Current evidence, history and uncertainty

temporal=auto uses explicit current/history cues; current requests current-state ordering and relevance keeps relevance ordering. Mixed current/history questions retain relevance ordering in auto mode. Occurrence questions preserve events instead of treating repeated evidence as a single fact. This is not an arbitrary-date state resolver.

Current-state comparison authenticates the source and checks the declaration's own field and environment. A heading, another field or a staging sentence cannot supply a missing production qualifier. Repeated identical declarations qualify only under conservative whole-source checks. Explicit no-change reports can yield to a comparable decision; a report's independently requested inspection or checklist fact remains evidence in its own right. Unknown prose, quotes and withdrawal language receive conservative treatment.

On the 62-question development diagnostic, each backend returns the expected first source and complete evidence for 38/38 answerable questions in each of four auto-mode settings (1,200/2,400 characters × one/five results). The separate 38-question diagnostic reaches 32/32 in those same settings. The recorded source is 4e4a8eb2; these are 496 and 304 response configurations per backend, not that many independent questions. 37 additional controls per backend check quotation, revocation, exact fields/scopes and preservation of independent report facts. Original failed attempts remain available. Responses and independent audit.

For assessed current requests, MCP compact output retains evidence_assessment, including potential-conflict page IDs. potential_conflict means comparable latest declarations in the inspected pool disagree; inspect their original pages. no_comparable_declaration means this bounded inspection established none. Neither status proves truth, authority or absence across the archive. Relevance/history, occurrence and skipped retrieval paths can remain not_assessed. Some pure-prose fields and unrecognized phrasing remain unsupported. Status semantics and measured limits.

Character and exact-token assembly preserve authenticated current evidence under the existing short-page rules. The final v0.13 source repeats 482 configurations per backend, preserving all 434 recorded current-evidence presences; the other 48 retain their original insufficient-budget outcomes. This is a correctness check, not 482 complete answers or a speed measurement. Final replay.

One project per MCP connection```bash

memory-book projects

Replace service-alpha with an exact ID returned above.

memory-book setup --client codex --mcp-project service-alpha
memory-book recall "current gradient accumulation steps" --project service-alpha --budget 2400 --limit 5


| Scope rule | Behavior |
|---|---|
| Bound connection, omitted project | Inherits the configured project |
| Conflicting project or foreign page ID | Rejected; cannot silently switch books |
| Ordinary client setup | Preserves an existing managed binding |
| `--mcp-project` / `--clear-mcp-project` | Explicitly change / remove the connection binding |
| `setup --project` | Capture mapping; requires `--capture` |

Unbound requests may search across books. A binding follows the configured connection, not the terminal's current directory; it protects against scope mistakes and is not multi-user authentication. Scope and source semantics.

## Source validation and delivery status

The published **v0.13.0** is built from `4977ebc13e9da10a8fa4cb95f39e2c533aabe790`. Release run 44343790035, attempt 1 completed all **30 native build/verification cells**—CPython 3.10–3.15 on five platforms—and its real public-URL installer. The original logs contain **4,320 native-test outcomes: 4,245 passes and 75 explicitly permitted optional skips**. The same source passed **11/11 jobs** in CI run 44343790028, attempt 1. Complete publication evidence.

The offline delivery audit verified **11 original artifact ZIPs and 67 independently downloaded public assets**. All **126 captured files** were then restored byte for byte; the original and restored audit JSONs match the independent proof exactly. This verifies saved delivery evidence without rebuilding or rerunning installation. Audit · Restoration proof.

The measured long-history source remains `e0e2515f6e16404cfb9b676cd2eab01ce84b81cc`, with the same complete Git tree as local `91412b43`; their identity receipt preserves both measured commit IDs. Its earlier CI run 44331775791, attempt 1 passed 11/11 jobs; that MCP job discovered 1,719 tests, passed 1,705 and skipped 14 with zero failures/errors. Original source-CI records.

**Historical v0.12.0 delivery** retains its separate publication proof: 11 successful CI jobs, 30 native cells, 4,245 native passes plus 75 permitted skips, 67 independently downloaded assets and successful public installation. These counts belong to v0.12. Historical v0.11 delivery.

Schema **4**, source/page IDs and Unicode coordinates remain unchanged. Use the Python interpreter printed by the installer in place of `python` to inspect the installed backend:```bash
python -I -X utf8 -c "import memory_book; from memory_book._text_backend import BACKEND_NAME; print(memory_book.__version__, BACKEND_NAME)"

On a compatible native installation, this prints 0.13.0 native; the portable fallback reports 0.13.0 python. MEMORY_BOOK_TEXT_BACKEND=auto is the default; python selects the reference, and native requires the extension. Backend and source-build options.

Documentation

Task Guide
Install, upgrade or run a service Deployment
Connect a client or enable capture Integration support matrix
Use citations, budgets and project scope Agent guide
Try engineering retrieval and backup Engineering workflow · Runnable example
Compare architectures and deployment needs Open-source comparison
Develop or release the project Contributing · Release runbook

Run memory-book serve to browse the archive at 127.0.0.1:8765. Archives are local and unencrypted; recalled evidence enters the selected client's model context. Data locations and maintenance.

Independently implemented under the MIT License.

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
AI Memory Book 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。

AI Memory Book 怎么在 Claude、Cursor、VS Code 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。

AI Memory Book 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。

AI Memory Book 需要 API Key 吗?

资料标注需要会话登录认证,使用前准备好对应凭据,并按最小权限配置。

AI Memory Book 支持哪些客户端?

资料标注支持 Claude、Cursor、VS Code、Windsurf、Gemini CLI、Codex;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。

AI Memory Book 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 2026-11-20。