FusionAL-Recall
MCP server for semantic search over a solved-issues engineering knowledge base. Built on FastMCP with sqlite-vec and sentence-transformers; exposes recall, remember, list_recent, verify, and get tools so agents can retrieve and log fixes.
README
FusionAL Recall — Semantic Registry Search MCP
Recall wraps the FusionAL Solved-Issues Registry with semantic search, three-tier access control, and auto-migration.
Query your team's solved problems via MCP tools: recall() for semantic search, remember() to log new entries, list_recent() for recent issues, verify() to validate entry integrity.
Features
- Semantic Search: Query by problem description, not keyword. "claude desktop timeout" returns SI-001 even if you don't say "timeout" or "claude" exactly.
- Automatic Migration: Parse existing SOLVED-ISSUES.md markdown on first run; auto-assign SI-XXX IDs; populate SQLite.
- Three-Tier Access:
personal(agent-only),project(team-scoped),public(shared community). - Vector Storage: sentence-transformers + sqlite-vec for fast CPU-friendly embeddings.
- MCP Tools:
recall(query, tier, limit),remember(symptoms, root_cause, fix, source, tags),list_recent(n),verify(si_id),get(si_id).
Architecture
SOLVED-ISSUES.md (markdown source)
↓
[Migrate on startup]
↓
SQLite + sqlite-vec (embeddings index)
↓
FastMCP Server (port 8107)
↓
[Tools]
SQL Schema
CREATE TABLE issues (
si_id TEXT PRIMARY KEY, -- SI-001, SI-002, etc.
title TEXT NOT NULL, -- one-line summary
symptoms TEXT NOT NULL,
root_cause TEXT NOT NULL,
fix TEXT NOT NULL,
source TEXT, -- session/task context
tags TEXT, -- comma-separated
verified_at TEXT, -- YYYY-MM
created_at TEXT NOT NULL, -- ISO 8601
tier TEXT DEFAULT 'personal', -- personal, project, public
embedding BLOB NOT NULL -- float32 vec (float-serialized)
);
Usage
Start the Server
python -m recall.server
# Listens on 0.0.0.0:8107
On first run, the server:
- Reads SOLVED-ISSUES.md (path from SOLVED_ISSUES_PATH env var)
- Parses SI-XXX markdown blocks
- Generates embeddings for symptoms + root_cause + fix (combined text)
- Inserts into SQLite + sqlite-vec index
- Auto-assigns next SI-ID based on highest existing ID
MCP Tools
recall(query: str, tier: str = "personal", limit: int = 5) → List[Issue]
Semantic search. Returns issues matching the query, ranked by embedding similarity.
# Query: "claude desktop timeout"
# Returns: SI-001 (Claude Desktop server timeout above 8 servers)
results = await client.call_tool(
name="recall",
arguments={"query": "claude desktop timeout", "tier": "personal", "limit": 5}
)
remember(symptoms: str, root_cause: str, fix: str, source: str, tags: str) → Dict
Log a new issue. Returns the assigned SI-ID (e.g., SI-012).
result = await client.call_tool(
name="remember",
arguments={
"symptoms": "My symptoms here",
"root_cause": "Root cause",
"fix": "Steps to fix",
"source": "session context",
"tags": "tag1,tag2"
}
)
# Returns: {"si_id": "SI-012", "title": "...", "created_at": "..."}
list_recent(n: int = 10) → List[Issue]
Return the N most recently added issues.
recent = await client.call_tool(
name="list_recent",
arguments={"n": 10}
)
verify(si_id: str) → Dict
Check if an SI entry exists and is valid.
valid = await client.call_tool(
name="verify",
arguments={"si_id": "SI-001"}
)
get(si_id: str) → Dict
Retrieve full entry by SI-ID.
entry = await client.call_tool(
name="get",
arguments={"si_id": "SI-001"}
)
Deployment
Docker
docker build -t fusional-recall:latest .
docker run -p 8107:8107 \
-e SOLVED_ISSUES_PATH=/mnt/kb/05-RECALL/SOLVED-ISSUES.md \
-v /path/to/kb:/mnt/kb \
fusional-recall:latest
Docker Compose
docker-compose up -d
# Recalls listens on localhost:8107
# recall.db volume persists embeddings across restarts
Testing
pytest tests/ -v
Tests validate:
- Migration: SOLVED-ISSUES.md parses correctly; all entries load into SQLite
- Recall: Semantic search returns SI-001 for query "claude desktop timeout"
- Remember: New entry is assigned SI-012, written to DB, queryable immediately
- Verify: SI-001 exists and validates; SI-999 does not
- Tier filtering: personal/project/public tiers are enforced
Environment Variables
| Variable | Default | Purpose |
|---|---|---|
RECALL_HOST |
0.0.0.0 |
MCP server bind address |
RECALL_PORT |
8107 |
MCP server port |
RECALL_DB_PATH |
./recall.db |
SQLite database file |
EMBEDDING_MODEL |
all-MiniLM-L6-v2 |
sentence-transformers model |
DEFAULT_TIER |
personal |
Default access tier |
SOLVED_ISSUES_PATH |
../fusional-knowledge-base/05-RECALL/SOLVED-ISSUES.md |
Path to registry markdown |
License
MIT. See LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。