MCP Knowledge Base Server
  
A Model Context Protocol (MCP) server providing persistent memory, knowledge base, and project summary capabilities with automatic project detection and an interactive dashboard.
How It Works
┌──────────────────────────────────────────────────────────────┐
│ AI Assistant (Kiro, Claude, etc.) │
│ │
│ "Remember we use JWT" "What do I know about auth?" │
│ "Init KB from my docs" "What changed since last summary?" │
└──────────────┬───────────────────────────┬───────────────────┘
│ MCP Protocol (stdio) │
▼ ▼
┌──────────────────────────────────────────────────────────────┐
│ mcp-kb-server │
│ │
│ ┌─────────┐ ┌──────────┐ ┌─────────┐ ┌──────────┐ │
│ │ memory │ │ kb │ │ summary │ │dashboard │ │
│ │ .store │ │ .add │ │.project │ │.projects │ │
│ │ .search │ │ .search │ │ .delta │ └──────────┘ │
│ │ .list │ │ .init ★ │ └─────────┘ │
│ │ .delete │ └──────────┘ │
│ │ .update │ Auto-detect project_id from project_root │
│ └─────────┘ package.json → git remote → directory name │
│ │
│ ┌────────────────────┐ ┌─────────────────────┐ │
│ │ memory.sqlite │ │ kb.sqlite │ │
│ │ + memory_fts(FTS5)│ │ + kb_fts (FTS5) │ │
│ │ + wiki_links │ │ + kb_meta (scoping) │ │
│ │ + expires/TTL │ │ + sources + fts │ │
│ └────────────────────┘ └─────────────────────┘ │
└──────────────────────────────────────────────────────────────┘
---
Workflow Diagrams
1. Request Flow
Every tool call follows the same path from AI to database and back.
AI Assistant
│
│ JSON-RPC 2.0 over stdio
▼
server.js
│
├─ Joi validation ──── invalid ──► error -32602
│
├─ project_id resolution
│ project_root provided?
│ ├─ yes → detectProjectId()
│ │ ├─ package.json name
│ │ ├─ git remote URL
│ │ └─ directory basename
│ └─ no → use explicit project_id
│
├─ LRU query cache (read-only tools)
│ hit ──────────────────────► return cached result
│ miss ─┐
│ ▼
├─ tool handler (memory / kb / sources / wiki / summary)
│ │
│ ▼
│ better-sqlite3 (synchronous)
│ memory.sqlite ── kb.sqlite
│ │
│ ▼
│ cache invalidated (on mutations)
│
└─ JSON-RPC result ──────────────► AI Assistant
---
2. Project Initialization (kb.init)
Run once when starting work on a project to populate the KB from existing docs.
Developer / AI
│
│ kb.init({ project_root: "/path/to/project" })
▼
Resolve project_id
│ package.json → git remote → directory name
▼
Expand glob patterns
│ default: ["**/*.md", "**/*.txt"]
│ skip: node_modules / .git / dist / build
▼
For each file
├─ already in KB? (matched by source path)
│ └─ yes + overwrite=false → skip
│
├─ read file content
│ └─ binary or empty → skip
│
└─ extract title
├─ first "# Heading" in content
└─ fallback: filename without extension
▼
Bulk INSERT (single SQLite transaction)
│ kb_fts ← title, content, source
│ kb_meta ← rowid, project_id
▼
Return { scanned, added, skipped, added_titles }
---
3. Memory Lifecycle
┌─────────────────────────────┐
│ AI Assistant │
└──┬──────────┬───────────┬───┘
│ │ │
store/update search delete
│ │ │
▼ ▼ ▼
┌────────────────────────────────────┐
│ memory.sqlite │
│ │
│ memory table │
│ ┌──────────────────────────────┐ │
│ │ id · project_id · scope │ │
│ │ content · tags · created_at │ │
│ │ updated_at · expires_at │ │
│ └──────────────────────────────┘ │
│ │ FTS5 sync │
│ memory_fts (BM25 ranked search) │
│ │
│ wiki_links (cross-references) │
└────────────────────────────────────┘
│
TTL purge
(auto on access, max 1×/60s/project)
│
▼
expired entries removed
---
4. Knowledge Base Flow
┌──────────────────────────────────┐
│ Input sources │
│ │
│ kb.init kb.add │
│ (bulk scan) (single doc) │
│ Dashboard Dashboard │
│ Import New Document │
└──────────┬───────────┬────────────┘
│ │
▼ ▼
┌──────────────────────────────────┐
│ kb.sqlite │
│ │
│ kb_fts ── title, content, source │
│ kb_meta ─ rowid → project_id │
└──────────────────────────────────┘
│
┌─────────────┼──────────────┐
│ │ │
kb.search dashboard wiki.link
(FTS5/BM25 (split-panel (cross-ref
+ Qdrant list+detail) to memory
optional) or source)
│
▼
AI Assistant
---
5. Summary & Delta Flow
summary.project summary.delta
────────────── ─────────────
project_root project_root
│ │
├─ auto-discover files ├─ run summary.project
│ README, ARCHITECTURE, │ (current snapshot)
│ .kiro/resources/*.md, etc. │
│ ├─ search memory for
├─ memory.search (recent) │ last "project-summary"
│ │ scope entry
├─ kb.search (relevant docs) │
│ └─ diff current vs stored
▼ │
combined snapshot text ▼
│ changed files / new memories /
│ (AI stores result via updated docs highlighted
│ memory.store scope=
│ "project-summary")
▼
searchable baseline for next delta
---
6. Dashboard Architecture
Browser HTTP Server (Node.js) SQLite
│ │ │
│ GET / │ │
│ ─────────────────────────► │ │
│ ◄── HTML (self-contained) ──│ │
│ │ │
│ GET /api/projects │ SELECT project_id, │
│ ─────────────────────────► │ COUNT(*) FROM memory ──► │
│ ◄── [{ project_id, total }] │ ◄─────────────────────── │
│ │ │
│ GET /api/kb │ SELECT kb_fts JOIN │
│ ?project_id=x&limit=30 │ kb_meta WHERE │
│ ─────────────────────────► │ project_id=x ──► │
│ ◄── { items, total } │ ◄────────────────────── │
│ │ │
│ (click item) │ │
│ GET /api/kb/:id ──► │ SELECT * FROM kb_fts ──► │
│ ◄── { title, content, … } │ ◄─────────────────────── │
│ │ │
│ GET /api/export/kb │ SELECT all docs for │
│ ?project_id=x ──► │ project → render .md ──► │
│ ◄── kb-x-2026-04-12.md │ ◄─────────────────────── │
│ │ │
│ POST /api/kb/import │ bulk INSERT kb_fts │
│ { project_id, docs[] } ──► │ + kb_meta (transaction)──►│
│ ◄── { ok, count } │ ◄─────────────────────── │
---
7. Wiki Links & Lint
Entries (any type)
memory ──┐
kb ──┼──► wiki.link ──► wiki_links table
source ──┘ (source_id, source_type,
target_id, target_type,
relation, project_id)
│
┌───────────────┼───────────────┐
│ │ │
wiki.links wiki.lint wiki.export
(lookup by (health check) (dump to .md)
entry_id) │
│ ┌────┴────────────────┐
│ │ orphan_memory │
│ │ (no links at all) │
│ │ │
│ │ broken_links │
│ │ (source or target │
│ │ no longer exists) │
│ │ │
│ │ stale_sources │
│ │ (90+ days, no links) │
│ └──────────────────────┘
▼
{ outbound[], inbound[] }
---
Features
🧠 Memory Management
- Long-term Memory: Store, search, list, update, and delete project-specific memory entries
- Full-Text Search: FTS5 with BM25 ranking via
use_ftsflag, plus substring fallback - Tag Filtering: Filter memories by tag name, combine with text queries
- Pagination:
memory.listwithtotal_count,offset,has_more - TTL / Expiry: Optional
expires_aton entries, auto-purged on access - Project Isolation: Complete data isolation between projects
📚 Knowledge Base
- Document Storage: Add and search documents with FTS5 full-text search
- Project Scoping:
project_idon all KB operations isolates docs per project - Bulk Init:
kb.initscans a project directory and imports all.md/.txtfiles in one call - Vector Search: Optional Qdrant integration for semantic search
- Source Tracking: Track document sources and metadata
📂 Sources Layer
- Raw Ingest: Ingest
.md,.txt,.csv,.pdffiles viasource.ingest - FTS Search: Full-text search across ingested source content
- Project Isolation: Sources scoped per project
🔗 Wiki Links & Lint
- Cross-references:
wiki.linkcreates typed links between memory/KB/source entries - Link Lookup:
wiki.linksreturns inbound and outbound links for any entry - Health Check:
wiki.lintdetects orphan entries, broken links, and stale sources - Export:
wiki.exportdumps all data as markdown files to a directory
📊 Project Summaries
- Snapshot Generation:
summary.projectreads files + memory + KB into a single snapshot - Delta Summaries:
summary.deltacompares current state vs. last summary - Auto-Discovery: Finds instruction files (README, ARCHITECTURE,
.kiro/resources/*.md, etc.)
🎨 Interactive Dashboard
- Split-panel UI: KB, Memory, and Sources tabs show a paginated list on the left and a detail view on the right — no more infinite scroll dumps
- Load More: All lists paginate in batches of 30
- Project Management: Create and delete projects directly from the UI
- KB Management: Add, edit, delete, import (JSON/
.md), and export KB documents as Markdown - Memory Management: View and delete individual memory entries; export all as Markdown
- Sources Viewer: Browse and preview ingested source files inline
- Dark / Light mode, responsive layout, markdown rendering
🔒 Safety
- Path validation, mismatch detection, XSS protection, parameterized SQL
Installation
npm install -g mcp-kb-server
Or use directly with npx:
npx mcp-kb-server
Usage
Configure MCP Client
{
"mcpServers": {
"kb-server": {
"command": "npx",
"args": ["mcp-kb-server"],
"env": {}
}
}
}
With a custom data directory:
{
"mcpServers": {
"kb-server": {
"command": "npx",
"args": ["mcp-kb-server"],
"env": {
"DATA_DIR": "/path/to/your/data",
"DASHBOARD_PORT": "4242"
}
}
}
}
---
Available Tools
Memory Tools
All memory tools require either project_root (recommended) or project_id.
memory.store ``json { "project_root": "/path/to/project", "content": "We use JWT for auth — stateless, works with mobile", "scope": "decisions", "tags": ["auth", "jwt", "architecture"], "expires_at": "2027-01-01T00:00:00Z" } ``
memory.search ``json { "project_root": "/path/to/project", "query": "authentication", "tag": "decision", "use_fts": true, "limit": 10 } ``
memory.list ``json { "project_root": "/path/to/project", "limit": 50, "offset": 0, "scope": "decisions" } ` Returns { total_count, offset, limit, has_more, entries }`.
memory.update ``json { "project_root": "/path/to/project", "id": "uuid", "content": "Updated content", "tags": ["updated", "auth"], "expires_at": "" } ``
memory.delete ``json { "project_root": "/path/to/project", "id": "uuid" } ``
---
Knowledge Base Tools
kb.add ``json { "title": "API Reference", "content": "## Endpoints\n...", "source": "docs/api.md", "project_id": "my-project" } ``
kb.search ``json { "query": "authentication", "project_id": "my-project", "limit": 5 } ``
kb.init ★ — Scan a project and bulk-import all docs into the KB ``json { "project_root": "/path/to/project", "patterns": ["*/.md", "docs/*/.txt"], "overwrite": false } ``
Returns: ``json { "project_id": "my-project", "scanned": 42, "added": 38, "skipped": 4, "added_titles": ["Getting Started", "API Reference", "..."], "skipped_paths": ["CHANGELOG.md (empty or binary)", "..."] } ``
- Auto-detects
project_idfromproject_root - Default patterns:
["*/.md", "*/.txt"] - Skips
node_modules,.git,dist,buildautomatically - Uses first
# Headingas title, falls back to filename - Skips already-imported files unless
overwrite: true - Bulk-inserts in a single SQLite transaction
---
Source Tools
source.ingest — Ingest a raw file (md, txt, csv, pdf) ``json { "project_root": "/path/to/project", "filename": "spec.md", "content": "...", "file_type": "md" } ``
source.list / source.search ``json { "project_root": "/path/to/project", "query": "authentication", "limit": 20, "offset": 0 } ``
---
Wiki Tools
wiki.link — Create a typed link between two entries ``json { "project_root": "/path/to/project", "source_id": "mem-uuid", "source_type": "memory", "target_id": "kb-rowid", "target_type": "kb", "relation": "implements" } ``
wiki.links — Look up all links for an entry ``json { "project_root": "/path/to/project", "entry_id": "mem-uuid" } ``
wiki.lint — Health check ``json { "project_root": "/path/to/project" } ` Returns { orphan_memory, broken_links, stale_sources, summary }`.
wiki.export — Export everything as markdown files ``json { "project_root": "/path/to/project", "output_dir": "./wiki-export" } ``
---
Summary Tools
summary.project ``json { "project_root": "/path/to/project", "auto_discover": true, "include_memory": true, "include_kb": true } ``
summary.delta ``json { "project_root": "/path/to/project", "auto_discover": true } ``
---
Dashboard Tool
dashboard.projects — Start the dashboard server and return the URL ``json { "limit": 10 } ``
Returns { dashboard_url, file_path }. Open the URL in a browser.
Set DASHBOARD_PORT=4242 in env to auto-start on server boot.
---
Dashboard
The dashboard is a self-contained HTML app served locally.
KB Tab
- Left panel: paginated document list (30/page) with search
- Right panel: full document view with Edit / Delete actions
- Toolbar: New Document, Import (JSON or
.md), Export as Markdown
Memory Tab
- Left panel: paginated entry list showing scope + preview
- Right panel: full markdown content with Delete button
- Header: Export all entries as Markdown
Sources Tab
- Left panel: paginated file list with search
- Right panel: inline file content preview (no modal)
Links / Lint Tabs
- Links: look up cross-references by entry ID
- Lint: detect orphans, broken links, stale sources
Project Management
- Project selector in topbar — switches all tabs
- "+" button to create a new project
- "Delete Project" button to wipe all data for the current project
Import Format (JSON)
[
{ "title": "Getting Started", "content": "## Overview\n...", "source": "docs/intro.md" },
{ "title": "API Reference", "content": "## Endpoints\n..." }
]
Export Format (Markdown)
KB export (kb-{project}-{date}.md): ```markdown
Knowledge Base — my-project
Exported 2026-04-12T... · 38 documents
---
Getting Started
Source: docs/intro.md ID: 1
Overview
... ```
Memory export (memory-{project}-{date}.md): ```markdown
Memory — my-project
Exported 2026-04-12T... · 12 entries
---
[decisions] uuid-here
Tags: auth, jwt Created: 2026-03-01T...
We use JWT for auth — stateless, works with mobile. ```
---
Automatic Project Detection
project_id is auto-detected from project_root in this order:
package.jsonnamefield- Git remote origin URL (repo name)
- Directory basename
All IDs are sanitized to lowercase alphanumeric + hyphens.
// Auto-detection
memory.store({ project_root: "/Users/me/my-app", content: "test" })
// → project_id: "my-app"
---
Environment Variables
| Variable | Default | Purpose | |---|---|---| | DATA_DIR | ./data | SQLite database location | | LOG_LEVEL | info | Winston log level | | MAX_MEMORY_ENTRIES | 1000 | Per-project memory cap | | MAX_KB_ENTRIES | 500 | Per-project KB cap | | ENABLE_QDRANT | false | Enable vector search | | QDRANT_URL | http://localhost:6333 | Qdrant endpoint | | VACUUM_INTERVAL | 86400000 | DB vacuum interval (ms) | | DASHBOARD_PORT | — | Auto-start dashboard on boot |
---
Project Structure
mcp-kb-server/
├── src/
│ ├── server.js # MCP entry point (JSON-RPC 2.0 over stdio)
│ ├── storage/db.js # SQLite schema + migrations
│ ├── tools/
│ │ ├── memory.js # memory.* tools
│ │ ├── kb.js # kb.add / kb.search / kb.init
│ │ ├── sources.js # source.ingest / list / search
│ │ ├── summary.js # summary.project
│ │ ├── summaryDelta.js # summary.delta
│ │ ├── wikiLinks.js # wiki.link / wiki.links
│ │ ├── wikiLint.js # wiki.lint
│ │ ├── wikiExport.js # wiki.export
│ │ └── dashboard.js # HTTP server + HTML dashboard
│ └── utils/
│ ├── projectId.js # Auto-detection logic
│ ├── fileDiscovery.js # Glob + file reading utilities
│ ├── config.js # Env config
│ ├── validation.js # Joi schemas
│ ├── performance.js # LRU cache + vacuum scheduler
│ └── errors.js # AppError / handleError
├── config/discovery.json # Auto-discovery glob patterns
├── data/ # SQLite databases (auto-created)
└── test/ # 84 tests (node --test)
---
Database Schema
memory.sqlite ``sql CREATE TABLE memory ( id TEXT PRIMARY KEY, project_id TEXT NOT NULL DEFAULT 'legacy', scope TEXT NOT NULL, content TEXT NOT NULL, tags TEXT, created_at TEXT NOT NULL, updated_at TEXT, expires_at TEXT ); CREATE VIRTUAL TABLE memory_fts USING fts5(content, tags); CREATE TABLE wiki_links ( id TEXT PRIMARY KEY, project_id TEXT NOT NULL, source_id TEXT, source_type TEXT, target_id TEXT, target_type TEXT, relation TEXT, created_at TEXT ); ``
kb.sqlite ``sql CREATE VIRTUAL TABLE kb_fts USING fts5(title, content, source); CREATE TABLE kb_meta (rowid INTEGER PRIMARY KEY, project_id TEXT); CREATE TABLE sources ( id TEXT PRIMARY KEY, project_id TEXT, slug TEXT, filename TEXT, file_type TEXT, content TEXT, file_path TEXT, ingested_at TEXT, size_bytes INTEGER ); CREATE VIRTUAL TABLE sources_fts USING fts5(slug, content); ``
---
Development
git clone https://github.com/dereknguyen269/mcp-kb-server.git
cd mcp-kb-server
npm install
npm test # runs all 84 tests
Run a single test file: ``bash NODE_ENV=test node --test test/memory.test.js ``
---
Security
- Project Isolation: SQL-level
project_idfiltering on every query - Path Validation: Prevents directory traversal
- XSS Protection: All user content HTML-escaped in dashboard
- Parameterized Queries: No SQL injection surface
- No External Resources: Dashboard works fully offline
---
Performance
- FTS5 / BM25: Ranked full-text search on memory, KB, and sources
- LRU Query Cache: 50-entry cache, 5-min TTL, invalidated on mutations
- Throttled Purge: Expired entry cleanup at most once per 60s per project
- SQLite WAL: Concurrent reads without blocking writes
- Bulk Transactions:
kb.initand dashboard import use single transactions
---
Error Codes
| Code | Meaning | |---|---| | -32602 | Invalid params (missing fields, bad paths) | | -32601 | Method not found | | -32000 | Server / DB error |
---
Version History
v1.2.0 (Current)
- ✅
kb.init— bulk-import project docs from filesystem in one MCP call - ✅ Dashboard split-panel UI — list + detail for KB, Memory, Sources
- ✅ Paginated lists (30/page, Load more) across all tabs
- ✅ Import KB from JSON array or
.mdfile via dashboard - ✅ Export KB and Memory as Markdown downloads
- ✅ Create / delete projects from dashboard
- ✅ Delete individual memory entries from dashboard
- ✅ KB list filtered by
project_id(was showing all projects) - ✅ New Document correctly scoped to current project
- ✅ 84 tests
v1.1.0
- ✅ memory.delete / memory.update / memory.list
- ✅ Tag-based filtering, FTS5 on memory
- ✅ KB project scoping, TTL/expiry, LRU cache
- ✅ Sources layer (ingest md/txt/csv/pdf)
- ✅ Wiki links, lint, export tools
v1.0.0
- ✅ Auto project_id detection
- ✅ Interactive HTML dashboard (dark/light)
- ✅ KB management (add/edit/delete)
- ✅ Project scoping, path validation
---
Support
- npm: https://www.npmjs.com/package/mcp-kb-server
- GitHub: https://github.com/dereknguyen269/mcp-kb-server
- Issues: https://github.com/dereknguyen269/mcp-kb-server/issues











