agoya
File-backed Memory MCP Server for multi-agent coordination, enabling AI coding agents to persistently store and search memories using JSON files without external services.
README
Oracle Memory
File-backed Memory MCP Server for multi-agent coordination.
Persistent memory layer for AI coding agents.
No database — just JSON files under .oracle-memory/.
Agents connect via stdio, HTTP, or streamable transport.
Requirements
- Node.js 24+
Quick start
npm install -g oracle-memory
# Run as MCP server (stdio)
oracle-memory
# Or with custom root:
ORACLE_MEMORY_ROOT_DIR=/path/to/project oracle-memory
Register in Claude Code:
claude mcp add oracle-memory -- /path/to/oracle-memory/dist/index.js
Register in Codex
Add to Codex MCP config pointing to the same path.
Memory types
| Type | Purpose | Lifetime |
|---|---|---|
fact |
Permanent knowledge (preferences, decisions, conventions) | Forever |
insight |
Lessons learned, gotchas, discoveries | Forever |
chunk |
Conversation snapshots (pre-compact) | Auto-expire (configurable TTL) |
working |
Session scratchpad, temporary context | Cleared between sessions |
MCP tools
| Tool | Description |
|---|---|
remember |
Save a fact/insight/chunk/working memory |
recall |
Search with hybrid BM25 + vector + entity graph ranking |
get_memory |
Retrieve a single memory by ID and type |
update_memory |
Update content/tags/importance/meta/TTL of an existing memory |
list_memories |
List memories with optional type/agent/tag/query filters |
forget |
Permanently delete a memory by ID and type |
clear_working |
Clear working memory for an agent (or all) |
consolidate |
Merge similar memories by tag overlap |
get_sessions |
List currently connected agent sessions |
get_stats |
Memory statistics by type and agent |
MCP resources
| URI | Content |
|---|---|
oracle-memory://memories |
All stored memories |
oracle-memory://memories/{type} |
Memories filtered by type |
oracle-memory://stats |
Memory statistics |
oracle-memory://sessions |
Currently connected agent sessions |
HTTP transport (multi-agent hub)
ORACLE_MEMORY_TRANSPORT=http ORACLE_MEMORY_PORT=8765 oracle-memory
Register each agent:
claude mcp add --transport http oracle-memory http://localhost:8765/mcp
With bearer auth:
ORACLE_MEMORY_HTTP_TOKEN=your-secret ORACLE_MEMORY_TRANSPORT=http ORACLE_MEMORY_PORT=8765 oracle-memory
Backward compatibility: Old
AGOYA_*env vars still work as fallbacks.
Storage layout
<root>/.oracle-memory/
├── config.json # Server configuration
├── facts/ # Permanent knowledge
├── insights/ # Lessons learned
├── chunks/ # Conversation snapshots
├── working/ # Session scratchpads
├── graph/graph.json # Entity relationship graph
└── vectors/ # Vector embeddings (optional)
All writes are atomic (write .tmp → rename). No corruption from crashes.
Search
BM25 keyword search (built-in, zero deps)
Tokenization + stop word filtering + BM25 ranking. Fast, deterministic, works offline.
Vector semantic search (optional)
When enabled, remember also indexes each memory with a vector embedding using
Xenova/all-MiniLM-L6-v2 (384-dim). On recall, results are fused using
RRF (Reciprocal Rank Fusion) — combining keyword relevance with semantic
similarity for the best of both worlds.
The model (~15MB) auto-downloads on first use and caches locally.
To disable:
ORACLE_MEMORY_DISABLE_VECTORS=1 oracle-memory
Example workflow
# Agent saves knowledge
→ remember(agent="claude", type="fact", content="Project uses port 3000", tags=["config"])
# Agent searches across sessions
→ recall(query="port configuration")
← [{ entry: { content: "Project uses port 3000", ... }, score: 2.3, method: "bm25" }]
# Update a memory
→ update_memory(id="20260713-...", type="fact", { content: "Project uses port 4000" })
# Check memory stats
→ get_stats()
← { totalMemories: 42, byType: { fact: 20, insight: 10, chunk: 10, working: 2 }, ... }
Environment
| Variable | Default | Description |
|---|---|---|
ORACLE_MEMORY_ROOT_DIR |
cwd |
Root directory for .oracle-memory/ store |
ORACLE_MEMORY_DISABLE_VECTORS |
false |
Set to 1 to disable vector search |
ORACLE_MEMORY_TRANSPORT |
stdio |
Transport: stdio or http/streamable |
ORACLE_MEMORY_HOST |
0.0.0.0 |
HTTP bind host |
ORACLE_MEMORY_PORT |
8765 |
HTTP port |
ORACLE_MEMORY_HTTP_TOKEN |
— | Bearer token for /mcp |
ORACLE_MEMORY_LOG_LEVEL |
info |
Log level |
Old AGOYA_* env vars still work as fallbacks.
Build
npm run build # TypeScript → dist/
npm run check # Type-check only
npm run dev # Run via tsx
npm start # Run compiled version
npm test # Run tests
Integrated projects
Oracle-Memory writes .oracle-memory/ format natively, making it interoperable with:
- Oracle — MCP-powered AI coding consultant (skills + oracles)
- Oracle Messages — Multi-agent message bus
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。
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。
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。
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。