Project Synapse
🧠 Project Synapse MCPサーバー
LLM-WIKI統合型 自律ナレッジ合成エンジン
Project Synapseは、Neo4j 2026.xグラフデータベースとObsidian Markdown Wikiを組み合わせ、永続的かつ蓄積型のナレッジベースを作成するMCP(Model Context Protocol)サーバーです。生のテキストはセマンティックパイプラインを通じて処理され、ベクトル埋め込みを持つ相互接続されたグラフノードへと変換されます。同時に、人間が読めるWikiレイヤーが、ブラウズ可能な相互リンクされたMarkdownページを提供します。
📚 ドキュメント
Project Synapseのセットアップと使用方法の詳細については、以下のガイドを参照してください:
Getting Started: インストールと初回実行。
Architecture: セマンティックパイプラインとグラフの仕組み。
Configuration: 環境変数の一覧。
Development: ツールの追加と貢献のためのガイド。
Testing: テストスイートの実行。
Contributing: プロジェクトへの貢献ガイドライン。
Related MCP server: KG Memory
これは何であり、何ではないのか
これはナレッジシステムであり、コードエディタではありません。 プロジェクトを取り巻く思考、研究、執筆のためのものです。例えば、アーキテクチャの決定、ドメイン調査、設計の根拠、参考資料、会議のメモなどが対象です。
コードはリポジトリに存在し、コードに「関する」知識はここに存在します。
ユースケース:
数週間から数ヶ月にわたって蓄積される詳細な調査
プロジェクトのナレッジベース(「何をしたか」だけでなく「なぜ決定したか」)
個人のナレッジマネジメント(記事、書籍、ポッドキャストのメモ)
AIをWikiの管理者とした共同ブレインストーミング
プロジェクトごとのセットアップ: プロジェクトごとに個別のObsidianボルトとGitHubリポジトリを作成してください。WIKI_VAULT_PATH環境変数をそこに指定します。1つのNeo4jインスタンスで複数のプロジェクトを運用可能です(グラフは共存します)。
アーキテクチャ
Web / Raw Sources
│
[defuddle] ← cleans web content before ingestion
│
▼
┌──────────────────────┐ ┌─────────────────────┐
│ Semantic Pipeline │────▶│ Neo4j Knowledge │
│ (Montague Grammar, │ │ Graph (entities, │
│ NLP, embeddings) │ │ facts, vectors) │
└──────────────────────┘ └────────┬────────────┘
│
┌───────┴───────┐
│ Wiki Adapter │
└───────┬───────┘
│
┌───────▼───────┐
│ Obsidian Vault│
│ (Markdown, │
│ Git-synced) │
└───────────────┘主な機能
ナレッジグラフ (Neo4j 2026.x)
ANNセマンティック検索を備えたネイティブな VECTOR 型
キーワード検索用のフルテキストBM25インデックス
ハイブリッド検索(ベクトル + BM25スコア融合)
隠れた関係性を発見するためのグラフ探索
形式意味論分析のためのモンタギュー文法パーサー
自律的な洞察生成のためのツェッテルカステンエンジン
LLM-WIKI統合
Obsidian MarkdownボルトとNeo4jグラフを橋渡し
YAMLフロントマター付きのフルページCRUD
自動インデックス生成と追記型ログ
ヘルスチェック:孤立ノードの検出、壊れたWikiリンク、不足しているフロントマター
効率的なグラフ同期のためのデルタ同期マニフェスト(コンテンツハッシュ化)
Webコンテンツの取り込み (defuddle)
wiki_fetch_urlは任意のURLを取得し、defuddleを介してナビゲーション/広告/不要な要素を除去し、Neo4jに取り込み、Clippings/にアーカイブします。1回の呼び出しで完全に自動化されます。wiki_ingest_rawはraw/から処理済みファイルをClippings/へ自動移動し、インボックスをクリーンに保ちます。raw/は真のインボックスであり、セッション終了後は空になります。
ローカル専用埋め込み (有料API不要)
sentence-transformers (デフォルト) — GPU上で実行
Ollama (オプション) — 任意のローカル埋め込みモデル
すべてのベクトルは
db.create.setNodeVectorProperty()を介してNeo4jにネイティブ保存
クイックスタート
前提条件
Python 3.12+
Neo4j 2026.x (CommunityまたはEnterprise)
uvパッケージマネージャー (
pip install uv)Wikiボルト用のGitHubリポジトリ(プライベート可)
Node.js + defuddle (Webコンテンツ取得用 — 下記参照)
Neo4jのセットアップ
# Ubuntu/Debian — see neo4j.com for other platforms
sudo apt install neo4j
sudo systemctl start neo4j
sudo systemctl enable neo4j
# Set password (default user: neo4j)
sudo neo4j-admin set-initial-password your_passworddefuddleのセットアップ
defuddleはWebページからナビゲーション、広告、定型文を除去し、クリーンなMarkdownを抽出します。wiki_fetch_urlに必要です。
# Install Node.js if not present (via nvm recommended)
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.39.0/install.sh | bash
nvm install --lts
nvm use --lts
# Install defuddle globally
npm install -g defuddle
# Verify
defuddle --version注意: Synapseは、シェルのPATHになくてもnvmパスを介してdefuddleを自動的に見つけます。
wiki_fetch_urlでdefuddleが見つからないと報告された場合は、nvmで管理されたNodeバージョンにインストールされていることを確認してください。
Obsidianボルトのセットアップ
Obsidianで新しいボルトを作成する(またはWikiリポジトリをクローンする)
Gitコミュニティプラグインをインストールする(設定 → コミュニティプラグイン → 閲覧 → "Git")
GitプラグインにGitHubの認証情報を設定する
ボルト構造(
raw/,wiki/,Clippings/,AGENTS.md)は初回実行時にSynapseによって自動作成されます
インストール
cd /home/ty/Repositories/ai_workspace
git clone <repository-url> project-synapse-mcp
cd project-synapse-mcp
uv venv --python 3.12 --seed
source .venv/bin/activate
uv add -e .
uv run python -m spacy download en_core_web_sm
cp .env.example .env # edit with your Neo4j password and vault path設定
.envを編集します:
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_password
NEO4J_DATABASE=neo4j
# Embedding — local only, no paid APIs
EMBEDDING_PROVIDER=sentence-transformers # or "ollama"
EMBEDDING_MODEL=sentence-transformers/all-mpnet-base-v2
EMBEDDING_DIMENSION=768
# Wiki vault
WIKI_VAULT_PATH=/path/to/your/obsidian-vault
WIKI_GITHUB_REPO=https://github.com/user/wiki-repoClaude Desktop / MCP統合
MCP設定に追加します:
{
"mcpServers": {
"project-synapse": {
"command": "uv",
"args": [
"--directory", "/path/to/project-synapse-mcp",
"run", "python", "-m", "synapse_mcp.server"
],
"env": {
"NEO4J_URI": "bolt://localhost:7687",
"NEO4J_USER": "neo4j",
"NEO4J_PASSWORD": "your_password",
"NEO4J_DATABASE": "neo4j",
"WIKI_VAULT_PATH": "/path/to/obsidian-vault",
"WIKI_GITHUB_REPO": "https://github.com/user/wiki-repo"
}
}
}
}MCPツール
ナレッジグラフ
ツール | 説明 |
| セマンティックパイプラインを通じてテキストを処理 → Neo4j |
| 洞察を優先した結果を返すベクトルセマンティック検索 |
| 隠れた関係性を発見するためのグラフ探索 |
| 自律的なツェッテルカステン・パターン検出 |
| モンタギュー文法による意味論的分析 |
Wiki (LLM-WIKI)
ツール | 説明 |
| URL取得 → defuddleでクリーン化 → 取り込み → Clippings/へアーカイブ |
| raw/からファイルを取り込み → Neo4j + Clippings/へ自動移動 |
| フロントマター付きWikiページの作成/更新 |
| パスによるWikiページの読み込み |
| Wikiページ全体のキーワード検索 |
| サブディレクトリ内の全ページをリストアップ |
| Wikiインデックスの再構築 |
| ヘルスチェック:孤立ページ、壊れたリンク、不足しているフロントマター |
Wikiボルトの構造
LLM-WIKI/
├── AGENTS.md # Agent schema doc — conventions and workflows
├── raw/ # INBOX ONLY — unprocessed files; empty after each session
├── raw-inbox.base # Obsidian Base view of pending raw/ queue
├── Clippings/ # Permanent archive — all processed sources land here
├── wiki/
│ ├── index.md # Auto-generated page catalogue
│ ├── log.md # Append-only activity log
│ ├── entities/ # People, tools, projects
│ ├── concepts/ # Ideas, theories, patterns
│ └── sources/ # Summaries of ingested sourcesコンテンツのライフサイクル
You clip/save → raw/ # your inbox
or
Agent fetches → wiki_fetch_url # web research
│
[defuddle clean]
│
[semantic pipeline] → Neo4j
│
wiki_write_page → wiki/sources/
│
auto-move → Clippings/ # permanent archiveraw/はセッション終了後に常に空になります。Clippings/は処理されたすべてのものの永続的な記録です。wiki/sources/内のソースページは、ファイルパスではなく元のURLを参照します。
ワークフロー
Webリサーチ:
wiki_fetch_url(url)→ 1回の呼び出しで取得、クリーン化、取り込み、アーカイブを実行手動クリップ:
raw/にドロップし、wiki_ingest_raw(filename)を呼び出す → 取り込み後に自動アーカイブクエリ:
query_knowledge(グラフ) またはwiki_search(ファイル) → 回答を合成Lint:
wiki_lint→ 孤立ページ、壊れたリンク、古い主張を修正ロールバック: Obsidian Gitプラグインを介してGitがバージョン管理を処理
理論的基盤
モンタギュー文法: 意味抽出のための形式的な合成意味論
ツェッテルカステン手法: 創発的な構造を持つアトミックなリンクノート
グラフ理論: コミュニティ検出、中心性、パス解析
Karpathy LLM-WIKI: ステートレスなRAGに対する永続的な知識コンパイル
ヴァネヴァー・ブッシュのMemex: 維持されたトレイルを持つプライベートな連想知識
ライセンス
MIT — LICENSEを参照してください。
Project Synapse: リアクティブなRAGから、永続的で蓄積型の知識へ。
Available Tools
23 toolsanalyze_semantic_structureB
Analyze the semantic structure of text using Montague Grammar parsing.
This tool provides insight into the formal semantic analysis capabilities and shows the logical form translations.
Args: text: Text to analyze semantically include_logical_form: Whether to include the formal logical representation
Returns: Semantic analysis with entities, relations, and optional logical forms
Inputs: text (required), include_logical_form. Returns: result
debug_testB
Simple test tool to check if MCP server is working.
Inputs: no named parameters. Returns: result
explore_connectionsB
Explore connections and relationships around a specific entity in the knowledge graph.
This tool implements the graph traversal capabilities for discovering non-obvious connections and patterns.
Args: entity: Entity name to explore from depth: How many relationship hops to explore (1-5) connection_types: Specific relationship types to follow
Returns: Visual representation of connections and discovered patterns
Inputs: entity (required), depth, connection_types. Returns: result
generate_insightsC
Trigger autonomous insight generation using the Zettelkasten engine.
This tool activates the autonomous synthesis engine to identify patterns and generate novel insights from the existing knowledge graph.
Args: topic: Optional topic to focus insight generation on confidence_threshold: Minimum confidence level for insights (0.0-1.0)
Returns: Generated insights with confidence scores and evidence trails
Inputs: topic, confidence_threshold. Returns: result
ingest_textB
Ingest and process text into the knowledge graph using semantic analysis.
This tool performs the core knowledge synthesis pipeline:
Semantic parsing using Montague Grammar
Entity extraction and relationship identification
Storage in the Neo4j knowledge graph
Automatic insight generation triggers
Args: text: Raw text to process and analyze source: Source identifier for provenance tracking metadata: Additional metadata about the text
Returns: Processing summary with entities and relationships extracted
Inputs: text (required), source, metadata. Returns: result
query_knowledgeA
Query the knowledge graph for facts and insights using natural language.
This tool provides the conversational interface to the knowledge base, prioritizing synthesized insights over raw facts.
Args: query: Natural language query include_insights: Whether to include AI-generated insights max_results: Maximum number of results to return
Returns: Query results with facts, insights, and reasoning trails
Inputs: query (required), max_results, include_insights. Returns: result
synapse_causal_windowA
Find candidate causes by temporal correlation.
Surfaces facts whose valid_from falls in the window
[before - within_days, before] and that share at least one
entity with facts about effect_entity.
This is exactly the "track everything you ate to find what caused the headaches" pattern — you record symptom onset, you record meals and medications, then this tool surfaces co-occurring events as candidates. The tool returns correlation; the human (or a downstream reasoning step) decides what caused what.
Args: effect_entity: The thing whose causes you're hunting (e.g. "headache", "rash", "build failure"). before: ISO date/datetime — when the effect was observed. within_days: How far back to search. Default 30.
Returns: Ranked list of candidate cause-effect pairings with day deltas.
Inputs: before (required), effect_entity (required), within_days. Returns: result
synapse_invalidateA
Mark a previously-recorded fact as no longer true.
Sets valid_to rather than deleting — the historical record stays
intact, but the fact is no longer "currently true" for default queries.
Args: subject/predicate/object: The triple to invalidate. ended: ISO date/datetime when the fact stopped being true. Defaults to now if omitted.
Returns: Number of facts affected.
Inputs: object (required), subject (required), predicate (required), ended. Returns: result
synapse_memory_statsA
Quick stats: how many temporal facts are stored, time span covered.
Inputs: no named parameters. Returns: result
synapse_recallA
Look up time-stamped facts about an entity.
Args:
entity: Name to look up. Matches subject, object, or both depending
on direction.
as_of: ISO date/datetime — if given, only facts valid at this point
in time are returned. Omit for "currently true" facts.
direction: "outgoing" (entity is the subject), "incoming" (entity
is the object), or "both" (default).
Returns: Newline-separated list of facts with timestamps. Empty if none found.
Inputs: entity (required), as_of, direction. Returns: result
synapse_rememberA
Record a time-stamped fact in Synapse's episodic memory.
Use this whenever something is worth remembering across sessions: decisions, observations, "Ty said X on date Y", health/diet/symptom log entries, project milestones.
Args: subject: Who or what the fact is about. Free-form name. predicate: The relationship verb. snake_case preferred (e.g. "started_taking", "moved_to", "decided_to_use"). object: The other side of the relation. valid_from: ISO date or datetime when the fact became true. If omitted, "now" is used. Bare dates → midnight UTC. valid_to: ISO date or datetime when the fact stopped being true. Omit for still-current facts. confidence: 0–1. Default 1.0 for explicit user statements; lower when the agent is inferring. source: Where this fact came from. Defaults to "agent:claude" for things Claude is recording. Use "user" or a filename for facts from explicit user statements or document ingestion. note: Free-form context. Stored in metadata for later recall.
Returns: The fact id (stable content hash — safe to call twice).
Inputs: object (required), subject (required), predicate (required), note, +4 more. Returns: result
synapse_timelineA
Chronological view of remembered facts.
Args: entity: Scope to one entity, or None for the global timeline. limit: Max number of rows. Default 50.
Returns: Time-ordered fact list, oldest first.
Inputs: limit, entity. Returns: result
wiki_cluster_pagesA
Cluster wiki pages by semantic similarity using GAAC (TF-IDF).
Identifies:
Natural topic clusters — pages that belong together
Missing links — same-cluster pages with no wikilink between them
Merge candidates — pages so similar they may be redundant (sim > 0.7)
Args: n_clusters: Number of clusters (auto = sqrt of page count if omitted).
Inputs: n_clusters. Returns: result
wiki_fetch_urlA
Fetch a URL with defuddle (clean markdown extraction), save to raw/, ingest into the knowledge graph, and archive to Clippings/.
Use this when researching the web — it strips navigation and clutter, leaving only the article content. Much cleaner than raw web_fetch.
Args: url: The URL to fetch and process. ingest: If True (default), immediately ingest into Neo4j after saving. Set False to save to raw/ only for manual review first.
Inputs: url (required), ingest. Returns: result
wiki_hits_analysisA
Compute HITS hub and authority scores on the wiki wikilink graph.
Authorities = pages cited by many others — load-bearing knowledge nodes. Hubs = pages that link to many good authorities — navigation layers.
Use to identify which pages need deepening (high authority) and which need comprehensive link coverage (high hub).
Inputs: no named parameters. Returns: result
wiki_ingest_rawA
Read a raw source file and ingest it into both the knowledge graph and wiki.
Reads from raw/, runs it through the Synapse semantic pipeline, stores in Neo4j, and creates a summary page in wiki/sources/.
Args: filename: Filename inside the raw/ directory.
Inputs: filename (required). Returns: result
wiki_lintA
Run a health check on the wiki vault.
Detects orphan pages, broken wikilinks, and missing frontmatter.
Inputs: no named parameters. Returns: result
wiki_list_pagesB
List all markdown pages in the wiki vault with pagination.
Args: subdir: Subdirectory to list ('wiki' or 'raw'). limit: Maximum pages to return (default 50, max 200). offset: Offset for pagination. tag: Filter by tag.
Inputs: tag, limit, offset, subdir. Returns: result
wiki_read_pageB
Read a wiki page.
Args: path: Relative path from vault root (e.g. 'wiki/concepts/rag.md'). mode: What to return: - "meta": metadata/frontmatter only (from index, 0 file reads) - "excerpt": metadata + first 500 characters of page body (from index) - "full": complete page content (reads file from disk)
Inputs: path (required), mode. Returns: result
wiki_searchA
Search wiki pages by keyword.
Args: query: Space-separated search terms. limit: Maximum results to return (default 10). subdir: Scope to subdirectory.
Inputs: query (required), limit, subdir. Returns: result
wiki_sync_indexA
Refresh the wiki index database from disk.
Runs automatically on server start and after write operations. Call manually if files were changed outside Synapse (e.g., Obsidian edits, git pull).
Inputs: no named parameters. Returns: result
wiki_update_indexB
Rebuild the wiki index from all wiki pages.
Args: deep: If True, performs a disk-level verification of all indexed files.
Inputs: deep. Returns: result
wiki_write_pageB
Write or update a wiki page with frontmatter.
Args: path: Relative path (e.g. 'wiki/entities/neo4j.md'). body: Markdown body content. summary: One-line summary for the index. tags: Comma-separated tags.
Inputs: body (required), path (required), tags, summary. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 23 tools
The tools are grouped into wiki, knowledge graph, and memory domains, but several have blurry boundaries: wiki_sync_index vs wiki_update_index, synapse_recall vs synapse_timeline, and ingest_text vs wiki_ingest_raw could cause misselection. Descriptions help clarify most overlaps, but not all.
Most tools follow verb_noun snake_case with clear wiki_ and synapse_ prefixes, but the core knowledge graph tools (ingest_text, query_knowledge, explore_connections) lack a shared prefix and debug_test breaks the pattern. The naming is mostly consistent but not fully systematic.
23 tools is in the heavy range and requires a lot of surface area to navigate, though each tool has a plausible role in the wiki/knowledge-graph/memory system. It is borderline: not overwhelming, but more than is clearly necessary.
The set covers wiki read/write/indexing, knowledge graph ingestion and querying, and episodic memory lifecycle management including invalidation and causal analysis. Minor gaps exist (no wiki page deletion, no explicit knowledge graph entity update/delete), but core workflows are well covered.
Maintenance
Related MCP Connectors
Personal knowledge base MCP server with semantic search, auto-categorization, metadata extraction
Markdown-based note-taking with a hosted MCP server. Your notes serve you and your AI.
Personal knowledge graph as an AI memory layer over MCP - read, save, and link your memories.
MCP server unifying ERPs, CRMs, APIs and knowledge base for Claude, ChatGPT and Gemini.
Related MCP Servers
- AlicenseAqualityDmaintenanceAn MCP server that transforms markdown notes into an AI-powered knowledge graph. It enables LLM clients to explore, analyze, and diagnose knowledge graphs through tools for node explanation, path finding, causal chain analysis, and wiki health reporting.5MIT
- FlicenseNot gradedqualityDmaintenanceAn MCP server that implements a heavily typed knowledge graph memory system with AI-powered entity and relation extraction, enabling structured knowledge storage and retrieval from unstructured text using predefined or custom ontologies.9-
- AlicenseNot gradedqualityDmaintenanceAn MCP server that compiles any text into a verifiable, graph-based knowledge base using deterministic chunking and parallel extraction of epistemology primitives.MIT
- AlicenseAqualityBmaintenanceAn MCP server that turns any compatible AI agent into a Zettelkasten partner for creating atomic notes, forming semantic links, detecting clusters, and synthesizing insights from your existing knowledge.192MIT