---
title: "Codex CLI"
description: "Connect your Codex CLI to CORE's memory system"
---

### Prerequisites

- [Codex CLI](https://codex.so) installed
- CORE account - [Sign up at core.heysol.ai](https://core.heysol.ai)

### Step 1: Configure CORE MCP Server

Create or open your Codex configuration file at `~/.codex/config.toml`:

```bash
# Create config directory if needed
mkdir -p ~/.codex

# Open config file in your editor
code ~/.codex/config.toml -r
```

### Step 2: Add CORE MCP Configuration

Add the following to your `config.toml` file:

```toml
[mcp_servers.corememory]
command = "npx"
args = ["-y", "mcp-remote", "https://core.heysol.ai/api/v1/mcp?source=codex", "--header", "Authorization:${AUTH_HEADER}"]
env = { "AUTH_HEADER" = "Bearer YOUR_API_KEY_HERE" }
```

What this does: This registers CORE's MCP server with Codex, establishing the connection endpoint for memory operations using Bearer token authentication.

### Step 3: Get Your API Key

1. Log into your CORE dashboard at [core.heysol.ai](https://core.heysol.ai)
2. Navigate to **Settings** (bottom left)
   ![CORE Settings](/images/core-settings.png)
3. Go to **API Key** → **Generate new key** → Name it "codex"
   ![Create API Key](/images/create-api-key.png)
4. Copy the generated key and replace `YOUR_API_KEY_HERE` in your config file

### Step 4: Verify Connection

Start Codex CLI and test your setup:

1. **Start Codex CLI** - The CORE memory MCP server should now be available
2. **Test memory storage**: Ask Codex to create a comprehensive summary of your codebase and add it to CORE memory for future reference
3. **Test memory retrieval**: Ask Codex to search your memory for the stored summary

## Enable Automatic Memory Integration (Recommended)

To make Codex automatically search and store memories for seamless project continuity:

Create `AGENTS.md` in your project root (if it doesn't exist, just append if it already exists):

```bash
touch AGENTS.md
```

Add the following to `AGENTS.md`:

```markdown
---
trigger: always_on
---

⚠️ **CRITICAL: READ THIS FIRST - MANDATORY MEMORY PROTOCOL** ⚠️

You are an AI coding assistant with access to CORE Memory - a persistent knowledge system that maintains project context, learnings, and continuity across all coding sessions.

## 🔴 MANDATORY STARTUP SEQUENCE - DO NOT SKIP 🔴

**BEFORE RESPONDING TO ANY USER MESSAGE, YOU MUST EXECUTE THESE TOOLS IN ORDER:**

### STEP 1 (REQUIRED): Search for Relevant Context

EXECUTE THIS TOOL FIRST:
`memory_search`

- Previous discussions about the current topic
- Related project decisions and implementations
- User preferences and work patterns
- Similar problems and their solutions

**Additional search triggers:**

- User mentions "previously", "before", "last time", or "we discussed"
- User references past work or project history
- Working on the CORE project (this repository)
- User asks about preferences, patterns, or past decisions
- Starting work on any feature or bug that might have history

**How to search effectively:**

- Write complete semantic queries, NOT keyword fragments
- Good: `"Manoj's preferences for API design and error handling"`
- Bad: `"manoj api preferences"`
- Ask: "What context am I missing that would help?"
- Consider: "What has the user told me before that I should remember?"

### Query Patterns for Memory Search

**Entity-Centric Queries** (Best for graph search):

- ✅ GOOD: `"Manoj's preferences for product positioning and messaging"`
- ✅ GOOD: `"CORE project authentication implementation decisions"`
- ❌ BAD: `"manoj product positioning"`
- Format: `[Person/Project] + [relationship/attribute] + [context]`

**Multi-Entity Relationship Queries** (Excellent for episode graph):

- ✅ GOOD: `"Manoj and Harshith discussions about BFS search implementation"`
- ✅ GOOD: `"relationship between entity extraction and recall quality in CORE"`
- ❌ BAD: `"manoj harshith bfs"`
- Format: `[Entity1] + [relationship type] + [Entity2] + [context]`

**Semantic Question Queries** (Good for vector search):

- ✅ GOOD: `"What causes BFS search to return empty results? What are the requirements for BFS traversal?"`
- ✅ GOOD: `"How does episode graph search improve recall quality compared to traditional search?"`
- ❌ BAD: `"bfs empty results"`
- Format: Complete natural questions with full context

**Concept Exploration Queries** (Good for BFS traversal):

- ✅ GOOD: `"concepts and ideas related to semantic relevance in knowledge graph search"`
- ✅ GOOD: `"topics connected to hop distance weighting and graph topology in BFS"`
- ❌ BAD: `"semantic relevance concepts"`
- Format: `[concept] + related/connected + [domain/context]`

**Temporal Queries** (Good for recent work):

- ✅ GOOD: `"recent changes to search implementation and reranking logic"`
- ✅ GOOD: `"latest discussions about entity extraction and semantic relevance"`
- ❌ BAD: `"recent search changes"`
- Format: `[temporal marker] + [specific topic] + [additional context]`

## 🔴 MANDATORY SHUTDOWN SEQUENCE - DO NOT SKIP 🔴

**AFTER FULLY RESPONDING TO THE USER, YOU MUST EXECUTE THIS TOOL:**

### FINAL STEP (REQUIRED): Store Conversation Memory

EXECUTE THIS TOOL LAST:
`memory_ingest`
Include the spaceId parameter using the ID from your initial memory_get_space call.

⚠️ **THIS IS NON-NEGOTIABLE** - You must ALWAYS store conversation context as your final action.

**What to capture in the message parameter:**

From User:

- Specific question, request, or problem statement
- Project context and situation provided
- What they're trying to accomplish
- Technical challenges or constraints mentioned

From Assistant:

- Detailed explanation of solution/approach taken
- Step-by-step processes and methodologies
- Technical concepts and principles explained
- Reasoning behind recommendations and decisions
- Alternative approaches discussed
- Problem-solving methodologies applied

**Include in storage:**

- All conceptual explanations and theory
- Technical discussions and analysis
- Problem-solving approaches and reasoning
- Decision rationale and trade-offs
- Implementation strategies (described conceptually)
- Learning insights and patterns

**Exclude from storage:**

- Code blocks and code snippets
- File contents or file listings
- Command examples or CLI commands
- Raw data or logs

**Quality check before storing:**

- Can someone quickly understand project context from memory alone?
- Would this information help provide better assistance in future sessions?
- Does stored context capture key decisions and reasoning?

---

## Summary: Your Mandatory Protocol

1. **FIRST ACTION**: Execute `memory_search` with semantic query about the user's request
2. **RESPOND**: Help the user with their request
3. **FINAL ACTION**: Execute `memory_ingest` with conversation summary and spaceId

**If you skip any of these steps, you are not following the project requirements.**
```

## How It Works

Once installed, CORE memory integrates seamlessly with Codex:

- **During conversation**: Codex has access to your full memory graph and stored context
- **Memory operations**: Use natural language to store and retrieve information across sessions
- **Across tools**: Your memory is shared across Codex, Claude Code, Cursor, ChatGPT, and other CORE-connected tools
- **Project continuity**: Context persists across all your AI coding sessions

## Troubleshooting

**Connection Issues:**

- Verify your API key is correct and hasn't expired
- Check that the `config.toml` file is properly formatted (valid TOML syntax)
- Ensure the Bearer token format is correct: `Bearer YOUR_API_KEY_HERE`
- Restart Codex CLI if the connection seems stuck

**API Key Issues:**

- Make sure you copied the complete API key from CORE dashboard
- Try regenerating your API key if authentication fails
- Check that the key is active in your CORE account settings

### Need Help?

Join our [Discord community](https://discord.gg/YGUZcvDjUa) and ask questions in the **#core-support** channel.

Our team and community members are ready to help you get the most out of CORE's memory capabilities.