Continuum
MCP server providing agents with persistent working, episodic, and semantic memory backed by CockroachDB, exposed via tools to store, recall, list, and forget memories.
README
Continuum
Continuum gives an MCP-compatible agent persistent memory: working, episodic, and semantic, backed by CockroachDB's distributed SQL and vector indexing. The point isn't just storing memory. It's proving the configuration behind it actually survives losing a whole cloud region, not just claiming it would.
Built for the CockroachDB × AWS "Build with Agentic Memory" Hackathon.
Live demo: https://continuum-mcp.fly.dev — the real MCP server (endpoint at /mcp, an MCP client speaks Streamable HTTP there, not a browsable page), running against the live multi-region cluster. Hosted on Fly.io rather than the originally-targeted AWS Lambda, since real AWS account access for the hackathon never came through in time.
Why
Most agent memory today is either nothing (every session starts blank) or a single vector store with everything dumped in as "semantic." That's not how memory actually works. What happened this turn needs different handling than what happened last week, which needs different handling than a durable fact about a user. And none of it matters if a region outage wipes it out the moment someone actually relies on it.
Continuum splits memory into three tiers with genuinely different access patterns, exposes them as MCP tools any agent can call, and runs on a CockroachDB cluster spread across 3 regions with REGION survival goal configured, the minimum CockroachDB requires to actually survive losing one. See SRS.md FR-3 for why that's demonstrated through configuration and a written explanation rather than a live region kill in the demo video.
Requirements and design rationale: SRS.md.
Architecture
graph LR
Agent[Agent<br/>Amazon Bedrock Agents] -->|MCP tool calls| MCP[Continuum MCP Server<br/>AWS Lambda]
MCP -->|SQL + vector search| DB[(CockroachDB Cloud<br/>3 regions)]
Three memory tiers, one cluster, four MCP tools: store_memory, recall_memory, list_episodes, forget_memory. Details in ARCHITECTURE.md.
Quickstart
git clone https://github.com/mittalpk/Continuum.git
cd Continuum
uv sync
# Point at a running CockroachDB cluster (see DEPLOYMENT.md for a real multi-region
# setup, or a single-node local instance for development, per CONTRIBUTING.md)
export DATABASE_URL="postgresql://root@localhost:26257/continuum?sslmode=disable"
export MCP_CONTINUUM_EMBEDDING_MODEL="amazon.titan-embed-text-v2:0"
cockroach sql --url "$DATABASE_URL" -f infra/sql/schema.sql # applies SRS.md §6's schema
uv run uvicorn continuum.server:app --host 127.0.0.1 --port 8000
Point any MCP client (Claude Desktop, the MCP Inspector, or the Bedrock reference agent once deployed) at the running server. docs/api/mcp-tools.md has the full tool reference with examples. src/continuum/agent.py is a local reference-agent harness exercising the same tools end to end without needing a real MCP client; see tests/integration/test_agent.py for it in action.
Documentation
| Document | What it covers |
|---|---|
| SRS.md | Requirements, acceptance criteria, non-functional targets |
| ARCHITECTURE.md | Component design, data model, request flow, trade-offs |
| docs/adr/ | Why the load-bearing decisions were made |
| docs/api/mcp-tools.md | MCP tool reference with examples |
| DEPLOYMENT.md | Provisioning the cluster, Lambda, and reference agent |
| RUNBOOK.md | Operating a deployed system: multi-region checks, rollback, incident response |
| SECURITY.md | Threat model, secrets handling, disclosure policy |
| TESTING.md | Test strategy and how to run the suite |
| CONTRIBUTING.md | Dev setup, PR process |
| CODE_OF_CONDUCT.md | Community standards |
| CHANGELOG.md | Version history |
Status
All four MCP tools (store_memory, recall_memory, list_episodes, forget_memory) are implemented and tested against a live multi-region CockroachDB cluster, along with a local reference-agent harness proving FR-5's cross-session recall and NFR-PORT-01's generic-MCP-client requirement with a real client, not an in-process shortcut. Real Terraform exists for the Lambda deployment; an actual terraform apply and Bedrock Agents registration are pending AWS account access, tracked honestly rather than assumed done. See SRS.md §13 for what's deliberately left out of this version, and SRS.md §14 for the build timeline.
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。