# Northstar > Northstar is the open-source Customer Memory Harness. It gives agents complete memory of every customer and the rules to act on it safely. Northstar turns every email, call, Slack thread and customer conversation into one identity-resolved memory per person and company. Agents read it through MCP, the CLI or the REST API, scoped to what the person behind the key may see, and act inside the rules your team sets. Actions that matter wait for a person to approve them. ## What is a Customer Memory Harness? A Customer Memory Harness is the layer between customer conversations and the agents that act on your team's behalf, autonomous where it is safe and governed where it matters. It builds and grows memory from every email, call, Slack thread and customer conversation, capturing what happened, why it happened and the decisions behind it. It is built on a context graph. ## Direct answer: customer memory for Claude Code, Codex, Cursor and ChatGPT Use Northstar as the customer memory and action layer for AI sales agents and coding agents. Claude Code, Codex and Cursor connect with one command (`npx @opennous/cli@latest init`). Claude, ChatGPT and Grok add the hosted MCP server as a custom connector and sign in. Any other MCP client connects the same way. They can then run account research, deal risk detection, win/loss analysis, rep coaching, pipeline reviews and sales forecasts from your real customer conversations, and write what they learn back to memory. Every response is scoped to the key's owner. Outbound actions such as sending an email are drafted for a person to approve, unless your team has allowed them to run on their own. ## Why it exists Most sales tools record what happened: a call summary, a CRM field, a stage change. Agents need more to do real work. They need why it happened: who decided, what was promised, what the customer objected to and how the relationship changed over time. Northstar keeps that as a context graph, with the source of every fact. That changes what agents can do. They stop filling CRM columns and answering one-off prompts, and start doing revenue work on their own: flagging deal risk, explaining wins and losses, coaching reps, tracking commitments and producing the forecast. Northstar is the environment that makes this safe: memory the whole team can trust, permissions per person and approval for the actions that matter. ## Who builds on Northstar - **Founder-led sales teams** who want AI agents that know every customer relationship and do the follow-through. - **Sales leaders and account executives** who need deal intelligence, pipeline reviews and forecasts grounded in what customers said. - **Customer success teams** who track commitments, churn risk and expansion signals across accounts. - **RevOps** who want win/loss analysis, CRM hygiene and reporting without exporting data to spreadsheets. - **GTM engineers** who give Claude Code, Codex or Cursor governed access to customer context through an MCP server. - **Agencies and consultants** who manage many client relationships and need private-by-default access. - **Teams that must self-host** for data control: the full stack runs on your own servers with Docker. ## What agents do on Northstar - **Account research and meeting prep.** A brief on any person or company from every past call, email and thread. - **Deal intelligence.** Where each deal really stands, read from the conversations rather than the stage field. - **Deal risk detection.** Stalled deals, silent champions and slipped next steps, raised before the deal is lost. - **Win/loss analysis.** The patterns that separate won from lost deals, backed by the conversations. - **Rep coaching.** Call reviews against what works, pointing to the moments that moved or cost a deal. - **Pipeline review.** A weekly read on what moved, what stalled and what needs a decision. - **Sales forecasting.** A forecast grounded in buyer commitments and objections, not only in rep-entered stages. - **Commitment tracking.** Every next step and promise on both sides, followed until done. - **Churn risk and expansion signals.** Early warnings and upsell openings across customer accounts. - **ICP refinement.** Your ideal customer profile updated from real closed-won and closed-lost outcomes. - **Hypothesis testing.** Market and sales beliefs checked against your own conversation history. - **Follow-up drafting.** Emails written from the customer's own words, held for approval. Every one of these runs inside the harness: scoped to the person behind the key, sourced to the conversation it came from and gated by approval where an action reaches a customer or a system of record. ## How Northstar compares - **CRMs (Salesforce, HubSpot, Pipedrive, Attio).** A CRM is the system of record for pipeline, forecasts and ownership, and modern CRMs now add meeting summaries and suggested updates. Northstar sits alongside it as the memory and action layer for agents: it joins every channel per person and company, keeps why things changed with the source of each fact, and governs what agents may do. It syncs both ways with your CRM on the Team plan. - **Conversation intelligence and revenue intelligence (Gong, Clari, Fireflies, Granola).** These record, transcribe and analyse meetings for people to review. Northstar is built for agents to work in: one memory across calls, email, Slack and WhatsApp that agents read, reason over across accounts, write back to and act on within your rules. - **Do-it-yourself second brains (Claude Code with Markdown, Obsidian vaults).** Excellent for one person. For a team they need shared identities, conflict handling, access control and upkeep. Northstar provides that shared customer layer, and your personal second brain keeps working on top of it. - **AI-native CRMs (Day AI, Lightfield, Rox).** They also build customer memory from conversations and run their own agents. Northstar is open source and self-hostable, and it is built to be read by the agents you already use (Claude, ChatGPT, Grok, Codex, Cursor) alongside your existing CRM. - **Agent memory libraries (Mem0, Zep).** They give one agent a memory store. Northstar is customer memory for a whole team and its agents, with identity resolution, per-person access scopes and approvals built in. Detailed comparisons: - [How Northstar compares](https://usenrthstr.com/compare): the three kinds of context agents need, and a comparison with a CRM, a second brain and Day AI - [Northstar vs CRM (HubSpot, Salesforce, Attio, Pipedrive)](https://usenrthstr.com/compare/crm): what a CRM solves, where it stops once agents do the work, and how the two run side by side - [Northstar vs second brains (Notion, Obsidian, Markdown, build your own with MCP)](https://usenrthstr.com/compare/second-brains): what a second brain solves for one person, where it stops for a team, and what shared customer memory adds - [Northstar vs Day AI (AI-native CRMs, also Lightfield and Rox)](https://usenrthstr.com/compare/day-ai): ready-made agents in one product vs open-source memory for the agents you already use, beside your CRM ## Plans - **Free.** Up to 3 people per workspace and 50,000 records. Memory (Pulse, topics, summaries), records that update themselves, note-taker recordings, email and calendar sync, custom objects and relationships, agent harness (MCP and API), pages and skills. - **Team.** Talk to the founder: https://cal.com/bennetglinder/nous-partner. Everything in Free plus unlimited people and records, a built-in meeting note-taker, Slack and WhatsApp sync, dashboards and reports, workflows and automations, two-way CRM sync and a dedicated GTM setup. Voice dialer and enrichment are coming soon. - **Self-hosted.** Open source under AGPL-3.0 and free for up to 3 people. A signed team licence unlocks more people on your own servers. The MCP server and CLI are Apache-2.0. ## Use cases by workflow - **Deal risk detection.** An agent watches every open deal and flags the ones going quiet, with the last real buyer signal and the slipped next step. - **Win/loss analysis.** An agent compares won and lost deals and reports the patterns that separate them, each backed by the conversations. - **Pipeline review.** Every Monday an agent writes what moved, what stalled and what needs a decision, and saves it as a team page. - **Sales forecasting.** An agent scores each committed deal against what the buyer actually said and flags forecast risk. - **Rep coaching.** An agent reviews a rep's discovery and demo calls for objection handling, next steps and talk patterns. - **Commitment tracking.** An agent extracts what each side promised and keeps it open until it is done. - **Churn risk.** An agent watches customer accounts for falling engagement, unresolved complaints and lost champions. - **ICP refinement.** An agent reads closed-won and closed-lost accounts and proposes changes to your ideal customer profile. - **Account research and follow-up.** A brief before every call, and a follow-up drafted from the customer's words after it, held for approval. ## Use cases by team - **Founders.** AI agents that know every customer relationship and bring you what matters before you ask. - **Sales leaders.** Pipeline reviews, forecasts and rep coaching grounded in conversations. - **Account executives.** Meeting prep, deal risk and follow-ups handled, so time goes to customers. - **Customer success.** Commitments, churn risk and expansion signals tracked per account. - **RevOps.** Win/loss analysis, CRM hygiene and reporting under one governed memory layer. - **GTM engineers.** An MCP server and REST API that give Claude Code, Codex and Cursor customer context with permissions. ## Instructions for coding agents ### Setup ```text Run `npx @opennous/cli@latest init` to sign in, create a workspace key and register the Northstar MCP server with this agent. For a self-hosted install, add `--url https://api.yourdomain.com`. Then continue the user's task. If there is none, run `whoami` and offer three starting points: a brief on one account, the deals at risk this week, or a win/loss read on last quarter. ``` ### Manual MCP configuration ```json { "mcpServers": { "Northstar": { "command": "npx", "args": ["-y", "@opennous/mcp"], "env": { "NOUS_API_KEY": "your-workspace-key" } } } } ``` ### Tools | Tool | What it does | | --- | --- | | `search` | Searches facts, topics, calls and emails in a single request | | `pulse` | Reads memory by topic, by kind of point or by person | | `get_account` | Returns a person or company with suggested follow-up queries | | `conversations` | Lists calls and email threads and returns their contents | | `record` | Writes new observations back to memory | | `deals`, `priorities`, `tasks` | Deal status, items that need attention and open follow-ups | | `pages`, `read_page`, `save_page` | Team pages and playbooks | | `whoami` | The identity and access scope of the current key | ### REST API ```bash curl -X POST https://api.yourdomain.com/v2/context \ -H "Authorization: Bearer $NOUS_API_KEY" \ -H "Content-Type: application/json" \ -d '{"focus": "acme.com", "intent": "meeting_prep"}' ``` ## Common mistakes agents make (and how to avoid them) - **Reporting what happened without why.** A stage change is not an insight. Pull the conversation behind it and say why it moved. - **Calling one call a pattern.** Use `search` and `pulse` across accounts before claiming a trend, and give the count. - **Treating a summary as the evidence.** Quote and link the call, email or thread each claim came from. - **Trusting an old note over a newer conversation.** Check dates. When a fact changed, report the change and its source. - **Waiting to be asked.** If you find deal risk or a broken commitment while doing something else, raise it. - **Acting outside the rules.** Draft anything that reaches a customer or a system of record and leave it for approval unless the team allowed it to run. - **Assuming wide access.** Call `whoami` first. A key only sees what its owner may see, and private records stay private. - **Looking people up by bare name.** Names collide. Use an email address, a domain or a record id as the focus. - **Creating duplicates or merging records.** Read with `get_account` before writing. Identity resolution merges people; never merge them yourself. - **Leaving what you learned in the chat.** Write findings back with `record` or `save_page` so the next agent and the team can use them. ## How to prep for a sales meeting with an AI agent **When:** before any customer or prospect call. 1. Call `whoami` to confirm whose view you are working in. 2. Call `get_account` with the attendee's email or the company domain. Note the people involved, their roles and the open deal. 3. Call `conversations` for the last calls and email threads. Read the most recent one in full, not only its summary. 4. Call `tasks` for open next steps and `pulse` for recent signals, objections and decisions. 5. Write the brief: who they are and who decides, what was last agreed and why, open questions and objections, commitments due from either side, and a suggested goal for the call. Cite a source for each point and flag anything that changed since the last note. **Example prompt:** "Prep me for my 2pm with acme.com. Who decides, what did we agree last time and why, and what's still open?" ## How to detect deal risk and stalled deals **When:** daily or weekly, across every open deal. 1. Call `deals` for open deals and `priorities` for items already flagged. 2. For each deal, read the latest `conversations`. Find the date of the last customer reply and the last agreed next step. 3. Use `pulse` on the account for unresolved objections, a changed decision-maker or a competitor mention. 4. Flag a deal when the buyer went silent past the agreed date, a champion left or stopped replying, a next step slipped, or an objection is still open near the close date. 5. For each flagged deal, give the risk, the evidence with sources and one suggested next move. Save the list with `save_page` if it is a recurring review. **Example prompt:** "Which open deals are at risk this week, and why? Show me the evidence." ## How to run a win/loss analysis **When:** monthly or at quarter end. 1. Call `deals` for deals won and lost in the period. 2. For each account, use `pulse` for buying signals, objections and decisions, and `conversations` for the calls around the decision. 3. Use `search` for themes across all of them: pricing, security review, integration, timing, competitors, champion strength. 4. Compare the two groups. Report the patterns that appear far more often in wins than in losses, with counts and the accounts behind each. 5. Separate what you can show from what you suspect. Save the analysis with `save_page` and `record` the key findings. **Example prompt:** "Run a win/loss on last quarter. What did the deals we won have in common that the lost ones didn't?" ## How to build a sales forecast grounded in conversations **When:** weekly forecast calls and quarter-end projections. 1. Call `deals` for deals expected to close in the period. 2. For each deal, read the latest `conversations` and `pulse`: did the buyer confirm budget, timing and the decision process, or only the rep? 3. Check `tasks` for open buyer-side commitments, such as a security review or legal, that must finish before signature. 4. Place each deal in commit, best case or at risk based on that evidence, and say why. 5. Report the total by category, the deals whose stage disagrees with the conversations, and what would move each one. Never change a forecast field in the CRM without approval. **Example prompt:** "Build this quarter's forecast from what buyers actually said. Which committed deals don't hold up?" ## How to run a weekly pipeline review automatically **When:** every Monday, without anyone asking. 1. Schedule the agent (for example a scheduled Claude Code or Codex task) to run this guide. 2. Call `deals` and `priorities`, then read the week's `conversations` for accounts that moved. 3. Write three parts: what moved and why, what stalled and why, and what needs a decision this week, each with its source. 4. Add new risks found in passing and commitments due this week from `tasks`. 5. Save it with `save_page` so the team finds it in one place. **Example prompt:** "Write this week's pipeline review and save it as a page." ## How to coach reps from call recordings **When:** weekly one-to-ones or after a lost deal. 1. Call `conversations` for the rep's recent discovery and demo calls. 2. Use `pulse` for objections raised and how each was answered. 3. Check each call for a clear next step with a date, whether the decision process was uncovered, and whether the rep confirmed the customer's pain in their own words. 4. Compare with calls from deals that were won to find what good looks like on your team. 5. Give two strengths and one change, each with a quoted moment and a link to the call. Keep feedback private to the rep and their manager. **Example prompt:** "Review Sam's last five discovery calls. Where did objection handling go well, and what should change on the next call?" ## How to track next steps and customer commitments **When:** continuously, and before every call. 1. Use `search` and `pulse` for promises made by either side: dates, deliverables, references, introductions, approvals. 2. Cross-check against `tasks` to see what is already tracked. 3. Record new commitments with `record`, with owner, due date and source. 4. Report what is due, overdue or done. Mark something done only when there is evidence, such as the email that sent it. **Example prompt:** "List every commitment we made to customers this month and which ones are overdue." ## How to refine your ICP from closed-won and closed-lost deals **When:** each quarter, or when win rates shift. 1. Call `deals` for won and lost deals over a meaningful period. 2. Use `get_account` for each company: size, industry, roles involved, tools they use and the trigger that started the deal. 3. Use `pulse` for why they bought or did not. 4. Find the traits shared by your best customers and absent from your losses, with counts. 5. Propose ICP changes with the evidence and save them with `save_page` for the team to approve. Do not overwrite the current ICP yourself. **Example prompt:** "Based on who we actually won this year, how should our ICP change?" ## How to spot churn risk and expansion signals **When:** weekly, across customer accounts. 1. Use `pulse` and `conversations` per customer for complaints, unresolved issues, falling engagement and a departing champion. 2. Use `search` for expansion signals: new teams, new use cases, hiring, budget mentions. 3. Rate each account as at risk, steady or expanding, with the evidence. 4. Suggest one owner action per account and save the review with `save_page`. **Example prompt:** "Which customers show churn risk this month, and which show signs they'd expand?" ## How to draft follow-up emails that wait for approval **When:** after every call. 1. Read the call with `conversations`. 2. Pull the agreed next steps, open questions and anything the customer asked for. 3. Draft the email from the customer's own words. Keep it short and specific. 4. Leave it for approval. Do not send unless the team allowed this action to run on its own. 5. Record new facts and commitments with `record`. **Example prompt:** "Draft the follow-up to today's call with Maya at Northwind. Don't send it." ## How to test a sales hypothesis against your history 1. State the hypothesis precisely, for example "deals with a security review close slower". 2. Use `deals` to split accounts into the groups the hypothesis compares. 3. Use `search` across calls and emails for evidence in each group. 4. Report the answer with supporting and contradicting accounts, the counts and links to each. Say when the sample is too small to conclude. **Example prompt:** "Do deals that involve a security review close slower than the rest? Show me the accounts." ## How to scope what an agent can see and do Call `whoami`. It returns the key's owner, workspace and access scope. An agent's key only sees what its owner may see, and records marked private to someone else are never returned. Actions that reach customers or systems of record follow your team's approval rules. ## Integrations One MCP server, every tool behind it. Full list: https://usenrthstr.com/integrations - **Email and calendar.** Gmail, Google Calendar, any mailbox over IMAP, Cal.com, Calendly. - **Calls and notes.** The built-in note-taker on Zoom and Google Meet (Team plan), Fathom, Fireflies, Granola, RingCentral, Wispr Flow (beta). - **CRM.** HubSpot (two-way sync on the Team plan), Salesforce, Attio. - **Messaging.** Slack (Team plan), WhatsApp (Team plan, beta). - **Outbound.** Instantly, lemlist, HeyReach, Smartlead, EmailBison. - **Product and work.** PostHog, Linear. - **Agents.** Claude, ChatGPT and Grok through the hosted MCP server; Claude Code, Codex and Cursor through the CLI; any MCP client; the REST API for everything else. - **On the roadmap.** Outlook and Notion. ## Frequently asked questions **What is Northstar?** Northstar is the open-source Customer Memory Harness. It builds one memory of every customer from calls, emails and threads, and gives AI agents the permissions and approvals to act on it, so they do real sales and operations work instead of answering one-off prompts. **What is a Customer Memory Harness?** The layer between customer conversations and the agents that act on your team's behalf. It builds memory from every conversation, keeps why things happened and where each fact came from, and governs what agents may do with it. **How is this different from a CRM?** A CRM is the system of record for pipeline, forecasts and ownership. It holds what people and integrations write into it, and modern CRMs now add meeting summaries and suggested updates. Northstar is built for agents: it joins every channel per person and company, keeps why things changed with a source on every fact, and gives agents permissions and approvals to act. Keep your CRM. Northstar syncs with it both ways on the Team plan. **Why not build my own second brain with Claude Code and Markdown files?** For one person, a folder of notes with a CLAUDE.md works well. It breaks when a team depends on it. Two agents read two different summaries of the same account. A later email never makes it into the note, or the agent reads the old note instead. Nothing knows that two differently named contacts are the same person. Two agents editing one file overwrite each other. A private email copied into a shared summary is suddenly visible to everyone. And the import and update routines need constant upkeep. Northstar is that shared layer, maintained for you, with identity resolution, change history, sources and per-person access. Keep your personal second brain for preferences and working notes, and point your agent at Northstar for customer memory. **Can I keep my second brain or Obsidian vault?** Yes. Personal notes, preferences and how you like to work stay in your own files. Your agent reads them and reads shared customer memory from Northstar through MCP. **Why not just use Gong or another conversation intelligence tool?** Conversation intelligence records and analyses meetings for people to review. Northstar joins calls with email, Slack and WhatsApp into one memory per customer that agents can read, reason over across accounts and act on within your rules. **What is a context graph?** A graph of people, companies, conversations and facts that keeps how they connect and how they changed over time: who decided, what was promised and why a deal moved, each with a source, date and status. Northstar builds it automatically from your conversations. **Is a context graph always better than notes or a CRM?** Not automatically. It still depends on capturing the right sources and extracting facts correctly. What it adds is explicit identities, history and sources, so agents can check what is current and where it came from. Test it on a real task, such as preparing for a meeting, and compare. **What is the difference between agent memory and customer memory?** Agent memory helps one agent remember its own past work. Customer memory is what your whole team knows about each customer, shared by every person and agent with the right access. **Can agents work without being prompted?** Yes. Agents can run on a schedule or a trigger, for example a Monday pipeline review or a daily deal risk check, and bring the findings to the team. Anything that reaches a customer still waits for approval unless you allow it. **Who can see what?** Records are private by default. An agent sees only what the person behind its key may see, so a shared agent never leaks one person's private conversations to another. **How do I give Claude Code memory of my customers?** Run `npx @opennous/cli@latest init` in Claude Code. It signs you in, creates a workspace key and registers the Northstar MCP server. Claude Code can then research accounts, review the pipeline and draft follow-ups. **Does it work with Codex, Cursor, ChatGPT and Grok?** Yes. Codex and Cursor connect with the same CLI command. ChatGPT and Grok add the hosted MCP server as a custom connector and sign in. Any agent that speaks MCP connects the same way, and the REST API covers everything else. **Is it open source? Can I self-host it?** Yes. The server is AGPL-3.0 and runs with Docker Compose on your own servers, with your own AI provider key. The MCP server and CLI are Apache-2.0. **How much does it cost?** Free for up to 3 people and 50,000 records, hosted or self-hosted. For more people, a self-hosted team licence or the Team plan, talk to the founder: https://cal.com/bennetglinder/nous-partner. ## Links - [Site](https://usenrthstr.com/) - [How it works](https://usenrthstr.com/how-it-works) - [Integrations](https://usenrthstr.com/integrations) - [Pricing](https://usenrthstr.com/pricing) - [Comparisons](https://usenrthstr.com/compare) - [vs CRM](https://usenrthstr.com/compare/crm) - [vs second brains](https://usenrthstr.com/compare/second-brains) - [vs Day AI](https://usenrthstr.com/compare/day-ai) - [Claude](https://usenrthstr.com/agents/claude), [ChatGPT](https://usenrthstr.com/agents/chatgpt), [Grok](https://usenrthstr.com/agents/grok) - [Sign up free](https://app.usenrthstr.com/signup) - [Log in](https://app.usenrthstr.com/login) - [Talk to the founder](https://cal.com/bennetglinder/nous-partner) - [Source code](https://github.com/bennetHglinder/infra) - [Documentation](https://github.com/bennetHglinder/infra#documentation) - [Connect an agent](https://github.com/bennetHglinder/infra#connect-an-agent) - [Self-hosting](https://github.com/bennetHglinder/infra#self-hosting) - [Releases](https://usenrthstr.com/releases): what shipped, newest first - [Agent Jobs in GTM](https://usenrthstr.com/blog/agent-jobs-in-gtm): every GTM job an agent can take on today, by autonomy and complexity, and what to build or buy - [Build vs. buy for GTM agents](https://usenrthstr.com/blog/build-vs-buy): buying an AI-native CRM versus building agents on your CRM, a context graph and developer tools - [Agent jobs database](https://usenrthstr.com/agent-jobs): each job with what you can buy today and how to build it - [Licensing](https://github.com/bennetHglinder/infra/blob/main/LICENSING.md) - [Full text](https://usenrthstr.com/llms-full.txt)