codex-substrate-memory
MCP server providing durable Substrate organizational memory tools, including search, read, query, ingest, remember, and sync. It integrates with Codex to automatically capture completed turns and session boundaries for persistent memory.
README
codex-substrate-memory
Public OpenAI Codex CLI plugin for durable Substrate organizational memory. It provides seven MCP tools, completed-turn capture through Codex's global notify hook, and a stale-rollout sweeper that emits content-free session boundaries.
Security requirement:
~/.codex/config.tomlis world-readable plaintext. The Substrate API key must come from the Codex process environment. Never putSUBSTRATE_API_KEYinconfig.toml, an[mcp_servers.substrate.env]table, a command argument, or this repository.
Install
Python 3.11 or newer is required.
python -m pip install codex-substrate-memory
For a source checkout:
python -m pip install -e .
Set the service URL and API key in the environment that launches Codex:
SUBSTRATE_API_URL=https://api.substrate.example
SUBSTRATE_API_KEY=<set-in-process-environment>
Wire Codex
Codex plugins do not declare MCP servers in .codex-plugin/plugin.json. MCP registration lives only in the user's ~/.codex/config.toml.
Codex also has only one global notify setting. If no notify hook exists, use:
notify = ["python", "-m", "codex_memory.notify"]
[mcp_servers.substrate]
command = "python"
args = ["-m", "codex_memory.server"]
startup_timeout_sec = 30
If a notify hook already exists, preserve it with the chainer. Replace the placeholders with the existing command and arguments in the same order:
notify = ["python", "-m", "codex_memory.notify", "--chain", "<original-exe>", "<original-arg-1>"]
[mcp_servers.substrate]
command = "python"
args = ["-m", "codex_memory.server"]
startup_timeout_sec = 30
substrate-codex configure reads the current TOML and prints the appropriate snippet. It is advisory and does not modify the file. substrate-codex configure --write is the explicit opt-in mutation path and creates a backup first.
Tools
| Tool | Purpose |
|---|---|
substrate_search |
Search organizational memory and return cited memory cards. |
substrate_read |
Read one wiki page by repository-relative path. |
substrate_query |
Ask a cited question across Substrate memory. |
substrate_ingest |
Submit intended text for asynchronous ingestion. |
substrate_remember |
Record a durable fact or decision. |
substrate_sync |
Sweep stale rollouts and flush the local spool. |
substrate_status |
Report content-free local configuration and delivery status. |
Capture model
Completed turns
Codex appends one JSON argument to the global notify command after an agent-turn-complete event. codex_memory.notify extracts input-messages, last-assistant-message, turn_id, and cwd; applies the shared bounded redaction and event builder; durably queues the turn; and attempts delivery. Capture failures never block Codex. With --chain, the original notify command receives the identical final JSON argument and its exit code is propagated.
Session end
Codex does not emit a session-end notification. codex_memory.rollout therefore scans ~/.codex/sessions/**/rollout-*.jsonl for rollouts older than 15 minutes. It tolerates truncated final lines, unreadable files, and unknown record types. For each stale rollout it queues one deterministic, content-free session_end event containing counts and a message-index boundary only, then persists a one-shot checkpoint marker. The sweeper runs from substrate_sync and substrate-codex sync.
Codex limitations
notifyis a single global setting. Installing this plugin must not silently replace another hook. Use the documented--chainform to preserve the original executable, arguments, final JSON argument, and exit code.notifyhas no session-end event. Session closure is inferred only after a rollout file has been inactive past the sweeper threshold. It is not instantaneous.
Privacy boundary
- API credentials are read from process environment variables only.
- Captured messages are normalized, redacted, bounded, and spooled before network delivery.
- The session-end sweeper sends no message text, rollout path, or working directory.
- Local state defaults to
~/.substrate/codex_memory;SUBSTRATE_STATE_HOMEoverrides the parent directory for isolated environments. - Redaction cannot recognize every sensitive statement. Use a trusted Substrate server and protect local state.
CLI
substrate-codex serve
substrate-codex status
substrate-codex sync
substrate-codex configure
python -m codex_memory.server is the direct MCP entry point.
Repository map
.codex-plugin/plugin.json— Codex plugin metadata; intentionally no MCP registration.skills/substrate-memory/SKILL.md— agent operating guidance.src/codex_memory/— Codex notify, rollout, configuration, CLI, and MCP integration.src/substrate_capture/— frozen shared capture core; do not edit.docs/config-snippet.toml— placeholder-only configuration example.tests/— shared-core and Codex integration tests.scripts/— vendor-manifest and public-plugin safety verification.
Verification scope
The repository test suite exercises the notify shape and chainer, tolerant rollout parsing, one-shot session-end emission, configuration safety, and an in-memory MCP initialize/list/call round trip. It does not claim a live Codex-to-production-Substrate end-to-end run.
License
MIT © 2026 Sightline Technologies Inc.
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。