buddhiAIClaude Code plugin
One install connects Claude Code to company memory and teaches it when to recall and when to propose.
The plugin is the quickest way to connect Claude Code to Buddhi. It sets up the connection, adds two commands, and reminds Claude to use company memory without anyone having to ask.
Install
In Claude Code:
/plugin marketplace add vanshobbyfi/buddhi_plugin
/plugin install buddhi-memory@buddhi
You are asked for two things:
| Setting | What to enter |
|---|---|
| Buddhi address | Your company's Buddhi address, with no trailing slash |
| Your Buddhi key | A key from Gateway & Agents. It starts with crtx_ |
Each person uses their own key. Claude then sees exactly what that person may see, and what it proposes is attributed to them.
Run /mcp to confirm plugin:buddhi-memory:buddhi is connected.
What you get
Recall without asking. Before answering a question about the company's decisions, products, customers or processes, Claude looks in company memory first, and says where the answer came from.
Proposals from real work. When a session produces something a colleague would need later — a decision and its reason, a fix that worked, what a customer asked for — Claude proposes it. A person approves every proposal before it becomes company memory. See Agent memory.
Two commands:
| Command | What it does |
|---|---|
/buddhi-memory:remember | Reviews the conversation and proposes what is worth keeping |
/buddhi-memory:remember <something> | Proposes that one thing |
/buddhi-memory:proposals | Shows which of your proposals are waiting, approved or rejected |
What runs automatically
- At the start of each session, a one-line reminder that company memory is connected.
- During a long session, an occasional check-in attached to one of your prompts, asking Claude to consider whether the work so far produced anything worth proposing. It adds no extra turn, and Claude stays quiet when nothing qualifies.
Turn the check-in off with Suggest proposals during long sessions under /config.
The plugin itself sends nothing to Buddhi. Only Claude's own tool calls do, and each one is visible in the session.
What Claude is told not to propose
- the conversation itself, or a summary of what it did
- anything already in company memory
- secrets, credentials, keys, or personal information about individuals
- temporary state such as current tasks
- code that already lives in the repository
Every proposal is read by a person, so a few good ones per session is the aim.
Requirements
- A current version of Claude Code.
- Node.js on your
PATH, for the two reminder scripts. Without it the connection and commands still work.
Other assistants
Any assistant that connects to Buddhi over MCP receives the same guidance when it connects: when to recall, what is worth proposing, and what never to submit. So Claude on the web, Claude Desktop, Cursor and others behave the same way without a plugin. The plugin adds the commands and the reminders. To connect those tools, see Connecting agents.