DecRec
Decision memory for AI agents: record, revisit, and resolve consequential choices.
Installation
Ajoutez ceci à la configuration MCP de votre client (Claude Desktop : claude_desktop_config.json ; Cursor : .cursor/mcp.json ; VS Code et les autres utilisent la même forme « mcpServers »).
claude mcp add --transport http decrec 'https://decrec.io/mcp'
{
"mcpServers": {
"decrec": {
"type": "http",
"url": "https://decrec.io/mcp"
}
}
}
Documentation
Tirée du README du projet sur GitHub, actualisée chaque semaine. Les liens ouvrent GitHub.
Sommaire
DecRec
AI explores. You decide. DecRec remembers.
Shared decision memory across Claude, ChatGPT, Cursor, Codex, and other agents.
Install · Docs · How it works · Why decisions · What the Skill does · What we learned · Privacy
Install in 30 seconds
This package is an Agent Plugin: a skill plus the DecRec MCP server.
In Cursor
Type this in the chat input as a slash command:
/add-plugin decrec@https://github.com/decrec-io/decrec-plugin
Or Settings → Customize → Plugins and install from that GitHub repo.
In Claude Code
/plugin marketplace add decrec-io/decrec-plugin
/plugin install decrec@decrec
In Codex CLI
codex plugin marketplace add decrec-io/decrec-plugin
codex plugin add decrec@decrec
Your agent forgot. DecRec didn't.
3 weeks ago — ChatGPT
Keep accounts single-user until we validate team demand.
✓ Committed as DEC-8
Today — Cursor, fresh conversation
Before changing the account model: DEC-8 deliberately kept accounts single-user while validating demand. Want to revisit it?
One decision. Any agent.
How it works
MCP gives the agent access to decisions. The Skill teaches it when those decisions matter.
Agent
├─ Skill → judgment: when decisions matter
└─ MCP → durable decision state
↕
DecRec
| Layer | Role |
|---|---|
| Skill | Judgment: what counts as a decision, when to recall, when to capture, when to stop. |
| MCP | Durable state: search, create, update, admit, commit, and archive decisions. |
| DecRec | The shared system of record across chats and agents. |
Tools without the skill still write; the skill without tools cannot persist across agents.
Why decisions, not everything?
DecRec remembers decisions, not AI output. Agents can prepare and surface records; a human commits the decision.
Your agent already has access to code, chats, docs, and project context. DecRec doesn't try to remember all of it.
It keeps the smaller set of choices that constrain future work:
- what you chose
- why
- what you considered
- who committed it
- what would justify revisiting it
The goal isn't more context. It's giving a fresh agent the reasoning behind decisions already made.
What the Skill does
- Stay in scope — don't advance past what was asked (no catalog → shape → update → explore → admit → commit chain unless requested).
- Recall — search only at a decision boundary (consequential choice that constrains later work). Never before ordinary coding.
- Capture — offer once when a consequential fork settles; never auto-create; never interrupt trivia.
- Humans commit — the agent can prepare and surface records, but it does not place commitments.
Full protocol: skills/decrec/SKILL.md. Exploration prompts: skills/decrec/explore-prompts.md.
Try it
After install, allow the MCP connection, then paste one of these into a new conversation:
Live choice
Should we [option A] or [option B]?
Work through it from more than one side, then record the decision in DecRec.
Or substitute any real decision you're working through.
A thread
Here is a [Slack thread / meeting notes]. What is worth recording in DecRec? Don't record until I pick.
[paste]
Already picked
We picked [option A] over [option B]. Record that decision in DecRec.
Install
| Client | Start here |
|---|---|
| Cursor | Slash command below, or Settings → Plugins |
| Claude Code | /plugin marketplace add |
| Codex CLI | codex plugin marketplace add |
| Claude (Web and Desktop) | Add marketplace decrec-io/decrec-plugin |
| ChatGPT | Work: marketplace plugin. Chat: MCP + skill zip |
| VS Code | Command Palette → Chat: Install Plugin From Source |
| Other | Copy skill files + MCP connector |
Cursor
Type this in the chat input as a slash command:
/add-plugin decrec@https://github.com/decrec-io/decrec-plugin
Or Settings → Customize → Plugins and install from that GitHub repo.
Claude Code
/plugin marketplace add decrec-io/decrec-plugin
/plugin install decrec@decrec
Codex CLI
codex plugin marketplace add decrec-io/decrec-plugin
codex plugin add decrec@decrec
Claude (Web and Desktop)
Same path on both. Skill and MCP come with the plugin.
- Customize → Plugins → Add → Add marketplace
- URL:
decrec-io/decrec-plugin - Sync → select DecRec → Install
- Allow when asked.
ChatGPT
Desktop (Work)
Marketplace plugin. Does not run in Chat.
- Open Plugins → Add → Add a marketplace.
- Source:
decrec-io/decrec-plugin - Add marketplace, then install decrec.
- Allow when asked. Start a new Work conversation.
Web and Desktop (Chat)
Marketplace plugin does not run here. Add MCP and the skill on chatgpt.com.
MCP (tools only — no skill):
- Plugins → +
- Connection: Server URL
- URL:
https://decrec.io/mcp - Create. Allow when asked. Start a new conversation.
Skill:
- Download ZIP from the GitHub repo
- Plugins → Skills → + → upload that zip
VS Code
Command Palette (Cmd+Shift+P / Ctrl+Shift+P), then:
Chat: Install Plugin From Source
Paste the GitHub URL and press Enter:
https://github.com/decrec-io/decrec-plugin
Other
Copy skills/decrec/*.md (SKILL.md and explore-prompts.md) into your client's skill directory. Add an MCP connector to https://decrec.io/mcp.
What we learned teaching agents when to remember
Useful even if you never use DecRec. Steal these for your own agent skills.
1. Capture only when all three hold
Alternatives (≥2 paths), consequence (constrains later work), future relevance (someone will ask why).
2. Don't advance past the ask
Match the stage they requested; stop there.
3. Recall only at a decision boundary
Not before ordinary coding, bugfixes, or implementing an already-chosen approach.
4. Fail-fast and branch on error codes
locked, in_flight, and pin_mismatch mean stop (or reload). Don't infer from the chat and retry.
5. Isolated role passes
Explore from clean contexts; the agent synthesizes options, the agent does not commit.
6. No invented facts
Don't invent stack, vendors, tenancy, or business metrics the brief doesn't state.
See skills/decrec/SKILL.md for the full rules.
Privacy
This plugin connects to the remote MCP server at https://decrec.io/mcp (OAuth). DecRec stores account email, organization membership, and decision content you choose to record. Details: Terms & data policy. Support: hello@decrec.io.
Questions fréquentes
Qu'est-ce que DecRec ?
Decision memory for AI agents: record, revisit, and resolve consequential choices.
Comment installer DecRec ?
Il fonctionne comme serveur distant. Dans Claude Code : claude mcp add --transport http decrec 'https://decrec.io/mcp'. Pour les autres clients, collez la configuration JSON de cette page dans leurs réglages MCP.
Faut-il une clé API pour DecRec ?
Son entrée au registre n'indique aucune clé ni secret obligatoire. Certaines fonctions peuvent tout de même demander un compte auprès du service auquel il se connecte.
DecRec est-il open source ?
Oui. Son code est sur GitHub sous licence MIT.
Qui maintient DecRec ?
Il est publié par decrec-io. AI.tn le référence sans le maintenir ni le recommander.
Issu du registre MCP officiel. Être listé n'est pas une recommandation : vérifiez le code et les permissions d'un serveur avant de lui donner accès à vos données.