Core – Personal AI OS with Memory Graph

TL;DR

TL;DR: Core is an open-source personal AI OS that maintains a persistent memory graph across every AI tool you use — indexes GitHub, Linear, Slack, and conversations; scores 88.24% SOTA on the LoCoMo benchmark; runs fully self-hosted.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: getcore.me ← MUST visit and verify
• Source repository: github.com/RedPlanetHQ/core ← MUST read README
• License: AGPL 3.0 ← MUST verify from LICENSE file
• HN launch thread: news.ycombinator.com/item?id=45093272 ← verified via HN Algolia API
• Docs: docs.getcore.me ← verified accessible
• Source last checked: 2026-08-01 (commit main branch)

What Is Core?

Core describes itself as “Your Personal AI OS” — not a chatbot you open, but an AI that is always running, always watching, and always context-aware. The core innovation is a temporal knowledge graph that remembers everything across every AI tool in your stack.

The product was verified against the GitHub README:

“CORE indexes email, meetings, GitHub, Linear, Slack and every assistant conversation into a temporal knowledge graph. Every task starts with the full picture — preferences, decisions, and prior context already loaded.”

Key Claims (from README)

| Feature | Details | |---|---| | Memory graph | Indexes 50+ apps via MCP into temporal knowledge graph | | Benchmark | 88.24% overall on LoCoMo (1,540 questions, 10 multi-turn conversations) | | Interfaces | Scratchpad, voice (Ctrl+Option on Mac), WhatsApp/Slack/Telegram messaging | | Coding agents | Delegates to Claude Code or Codex with full task context | | Browser | Built-in isolated browser CORE drives on your behalf | | License | AGPL 3.0 | | Stars | 1,935 on GitHub (verified via GitHub API) |

LoCoMo Benchmark Scores (from README)

• Single-hop recall: 91%
• Multi-hop reasoning: 85%
• Temporal understanding: 88%
• Open-domain synthesis: 71%
• Overall: 88.24%

Setup Workflow

Prerequisites

• Docker 20.10+ and Docker Compose 2.20+
• 4 vCPU / 8GB RAM
• Node.js (for the CLI install method)

Step 1: Install Core

npm install -g @redplanethq/corebrain && corebrain setup

The setup wizard asks for:

• Install directory
• AI provider and API key
• Chat model selection

It generates secrets, starts the Docker stack, and opens http://localhost:3033.

Step 2: Connect the Gateway

The gateway enables CORE to run coding agents, drive the browser, and access local folders:

corebrain login
corebrain gateway setup

Step 3: Connect Your First App

With the gateway running, connect GitHub, Linear, Slack, or any of the 50+ supported apps via MCP integrations at http://localhost:3033.

Step 4: Post Your First Task

Open the Scratchpad at http://localhost:3033 and type a task:

[ ] Summarize my open GitHub issues

CORE picks it up within 3 minutes, gathers context from connected apps and memory, and drafts a plan for your approval before executing.

Alternative: Railway Deploy

One-click deploy to Railway for server/VPS hosting:

Deeper Analysis

How the Memory Graph Works

Core’s memory graph is a temporal knowledge graph — every piece of information is stored with timestamps and relationships. When you switch between ChatGPT, Cursor, and Claude Code, Core loads the full context from previous sessions automatically.

From the README, the architecture supports:

• Subscription proxy — point Core at your existing Claude Pro or Codex subscription instead of paying per-token. Memory, planning, task execution, and coding sessions all use your subscription.
• MCP integrations — 50+ one-click connectors for GitHub, Linear, Jira, Slack, Gmail, Calendar, Sentry, Notion, Todoist, and more.
• Webhook triggers — inbound events can automatically create tasks.

Comparison with Related Tools

| | Core | OpenClaw | Devin / Copilot | |---|:---:|:---:|:---:| | Multiple interfaces (voice, scratchpad, chat, messaging) | ✅ | Partial | ❌ | | Persistent memory across tasks | ✅ | ❌ | ❌ | | Delegates to coding agents (Claude Code, Codex) | ✅ | ❌ | ✅ | | Structured task planning with human approval | ✅ | ❌ | Partial | | Custom name, personality, and voice | ✅ | ❌ | ❌ | | 50+ app connectors | ✅ | Partial | ❌ | | Open source and self-hostable | ✅ | ✅ | ❌ |

Source: Core README comparison table

Self-Hosting Requirements

Core self-hosting requires Docker 20.10+, Docker Compose 2.20+, and 4 vCPU / 8GB RAM minimum. The full stack includes a web UI, backend services, and the gateway component for browser/terminal access.

Practical Evaluation Checklist

• [ ] Docker and Docker Compose installed before running corebrain setup
• [ ] corebrain setup completes without errors
• [ ] Web UI accessible at http://localhost:3033
• [ ] Gateway setup (corebrain gateway setup) succeeds
• [ ] At least one app connected (GitHub or Slack recommended)
• [ ] Task submitted via Scratchpad resolves with context loaded
• [ ] Memory persists across restarts
• [ ] Voice input (Ctrl+Option on Mac) functional

Security Notes

• All data stays in your own infrastructure when self-hosted.
• The subscription proxy option uses your own Claude Pro or Codex API key — no separate key management required.
• The isolated browser component means agent browser sessions don’t share cookies with your normal browser.
• AGPL 3.0 license requires that modifications to Core itself be open-sourced if distributed.

FAQ

Q: What is the LoCoMo benchmark? A: LoCoMo is a rigorous benchmark for AI memory systems involving 1,540 questions across 10 multi-turn conversations, testing single-hop recall, multi-hop reasoning, temporal understanding, and open-domain synthesis. Core achieved 88.24% overall on this benchmark.

Q: How is Core different from a vector database or MEMORY.md file? A: Unlike flat memory files or vector databases, Core uses a temporal knowledge graph — raw conversation episodes get distilled into structured, time-stamped knowledge automatically, similar to how human memory works.

Q: Can I use Core without a Claude or Codex subscription? A: Yes. Core can use any OpenAI-compatible API key directly. It also supports pointing at an existing Claude Pro or Codex subscription via a subscription proxy.

Q: What messaging platforms does Core support? A: Core supports WhatsApp, Slack, and Telegram as messaging interfaces — you can send a task from any of these and CORE will work on it with full context from your memory graph.

Q: Is there a hosted option or does it have to be self-hosted? A: Core is open source and self-hosted only. A Mac app is on a waitlist at getcore.me, but no hosted SaaS version currently exists.

Conclusion

Core fills a real gap in the AI tooling stack: every new AI tool requires re-explaining your project context, decisions, and preferences. With Core’s memory graph, that context persists across ChatGPT, Cursor, Claude Code, and any other MCP-compatible tool.

The benchmark numbers (88.24% on LoCoMo) and the breadth of integrations (50+ apps) make it the most capable open-source memory solution available. The AGPL 3.0 license keeps it open, and the Docker-based install makes self-hosting straightforward for developers already familiar with container tooling.

If you run multiple AI coding agents or context-heavy AI tools, Core is worth integrating into your stack to eliminate the repeated context tax on every new session.

• Core on GitHub
• Core Documentation
• HN Discussion