CognitiveX memory for Claude Code

WHAT CLAUDE CODE FORGETS

Claude Code resets between sessions. The LCM doesn't.

Add one MCP server and the agent gets a memory that survives the window.

Claude Code is excellent at reasoning over the code in front of it. What it doesn't carry on its own is why you made a decision three sessions ago, which library you standardized on, or the naming convention you agreed in another repo. A hand-kept CLAUDE.md captures some of that, but it lives in one project and you have to remember to update it, and long sessions get compacted, which trims the context the agent was relying on.

CognitiveX fills exactly that gap. It's the Large Cognition Model, a cognitive layer that sits beside any LLM and gives it a persistent, evolving memory. You connect it to Claude Code through the Model Context Protocol (MCP), and from then on the agent can remember and recall across every session, every project, and every restart. The memory is the model, and it lives outside the terminal, so closing the session doesn't erase it.

SIXTY-SECOND SETUP

Wire CognitiveX into Claude Code over MCP

Claude Code speaks MCP natively, so there's nothing to fork or patch. Install the CLI, point it at your API key, and the remember / recall tools light up.

1. 1

   Install the CLI and connect Claude Code

   One command installs the server and registers it with Claude Code:

   # install the CognitiveX CLInpm i -g @cognitivx/cli # sign in, then register the MCP server with Claude Codecogx auth logincogx mcp install claude

2. 2

   Or add it to .mcp.json by hand

   Prefer to wire it manually? The endpoint is hosted, so there's nothing to run locally. Add it to your project's .mcp.json (or your user config for all projects):

   {  "mcpServers": {    "cognitivx": {      "url": "https://api.cognitivx.io/mcp",      "headers": {        "Authorization": "Bearer YOUR_COGNITIVX_KEY"      }    }  }}

3. 3

   Confirm the tools, then teach it once

   Run /mcp in Claude Code and you should see the CognitiveX tools light up: remember, recall, reflect, and the rest. Then tell the agent something worth keeping, like “remember we chose Drizzle over Prisma for the edge runtime”. Open a fresh session tomorrow, ask “what's our ORM decision and why?”, and Claude Code recalls it through CognitiveX. The context you build is no longer trapped in one conversation.

Want to wire the same memory into your own tooling? It's reachable from the cogx SDK and HTTP API too. See the developer docs and the platform overview.

FOUR TIERS, NOT A NOTES FILE

What Claude Code actually keeps.

CognitiveX isn't a flat log the model greps. It's a structured, four-tier memory, and Claude Code writes into whichever tier fits as you work:

• Decisions (semantic): the ORM you standardized on, the API you settled, why a tradeoff went the way it did.
• Sessions (episodic): what you shipped Tuesday, the bug you traced to a race condition, the migration you ran.
• Conventions (procedural): your test command, your branch naming, the way this repo handles config and secrets.
• Preferences (foundational): your stack tastes, tone, and the rules you never want re-litigated.

Episodic events get promoted into durable semantic facts, patterns surface on their own, and an overnight dream-consolidation pass reorganizes the substrate so recall stays sharp instead of drowning in stale notes. The memory gets better the more you use it. Read how the loop works on the LCM page.

NOT JUST CLAUDE CODE

The same memory follows you everywhere.

Because CognitiveX connects over MCP (and through the cogx SDK and HTTP API), the memory you build in Claude Code isn't locked to Claude Code. Open Cursor for an edit, ask Codex or ChatGPT a question, or hit the API from a CI job: they all read and write the same living memory. Decide something in the terminal in the morning and your editor already knows it in the afternoon.

• cross-session: survives new sessions, /compact, and machine restarts
• cross-tool: one memory shared across every MCP client you use
• cross-repo: conventions and decisions aren't trapped in one project folder
• storing is free: you only spend recall credits when you read memory back

See the full picture on the integrations overview, or read why shared memory across AI tools changes how you work.

Give Claude Code a memory.