Skip to content

x-community-post.md

Latest commit

 

History

History
77 lines (53 loc) · 3.08 KB

x-community-post.md

File metadata and controls

77 lines (53 loc) · 3.08 KB

How I Built Persistent Memory for Claude Code Agents (and Why Your AI Forgetting Everything Is a Solved Problem)

Every Claude Code session starts blank. Your agent forgets your name, your codebase conventions, the decision you made 20 minutes ago. You end up re-explaining the same context over and over.

I got tired of this and built Cortex — a local-first memory engine that gives AI agents real long-term memory. Here's what it does and why I think it matters.

The Problem

Current "memory" for AI agents is one of:

  • Flat text files — grep-based, no structure, no decay, no relationships
  • Cloud APIs (Mem0, etc.) — 200-500ms latency per query, $99+/mo, your data on someone else's servers
  • OpenAI Memory — opaque, no export, no control

None of these give you what human memory actually does: structured recall, belief updating, relationship tracking, and forgetting things that don't matter anymore.

What Cortex Does Differently

Cortex is a pure Rust MCP server (3.8MB binary, zero runtime deps) that runs 100% locally:

  • 4 memory tiers — Working → Episodic → Semantic → Procedural (like human memory)
  • Bayesian beliefs that self-correct with new evidence
  • People graph with cross-platform identity resolution
  • Sub-millisecond everything — 156µs ingest, 568µs search (528x faster than Mem0)
  • 25 MCP tools — plug into Claude Code, Claude Desktop, or any MCP client
  • AES-256-GCM encrypted sync through your own cloud storage (iCloud/GDrive/OneDrive/Dropbox)
  • Zero telemetry, zero cost, forever

Real-World Example: Claude Code + Cortex

I use Cortex with Claude Code for my X-Auto project (automated social media + SEO). Here's what changes:

Before Cortex: Every session I had to re-explain that the project uses Gemini as its LLM provider, that we push directly to main, that tests must pass before commits, and dozens of project-specific details.

After Cortex: Claude Code remembers all of this across sessions. It knows my code style preferences, which modules I've been working on, what decisions we made last week and why. The memory_context tool gives it a token-budgeted summary of everything relevant to the current task.

Benchmark Results

We tested against the LoCoMo benchmark (ACL 2024) — 1540 QA pairs across long-term conversations:

System Overall
Cortex v1.7 73.7%
Mem0-Graph 68.4%
Mem0 66.9%
OpenAI Memory 52.9%

Cortex beats Mem0 by 7 percentage points while running entirely on your machine with zero API costs.

How to Try It

# Install (macOS/Linux)
curl -fsSL https://github.com/ghraw/gambletan/cortex/main/install.sh | bash

# Or with npm
npx cortex-mcp

Add to your Claude Code MCP config:

{
  "mcpServers": {
    "cortex": {
      "command": "cortex-mcp-server",
      "args": ["--db-path", "~/.cortex/memory.db"]
    }
  }
}

That's it. Claude Code now has persistent memory.

Open Source, MIT Licensed

Full source: https://github.com/gambletan/cortex

If it's useful, a star helps others find it. Questions welcome.