World Model MCP: Memory Layer That Cut SWE-bench Repeat Mistakes by +10.2 Points

World Model MCP adds a temporal knowledge graph to Claude Code that learns from corrections, prevents repeated mistakes, and re-injects context after compaction — proven with +10.2 pts on SWE-bench.

What Changed

World Model MCP (v0.9.1) is a new MCP server that gives Claude Code long-term memory. It creates a temporal knowledge graph of your codebase that learns from every coding session. The key claim: it reduces repeated mistakes by +10.2 points on SWE-bench Verified.

The repo ships 26 MCP tools, 19 CLI subcommands, and 375 tests. It's harness-neutral — works with Claude Code, Cursor, and pi.

What It Does

World Model MCP acts as a persistent memory layer that:

1. Prevents Hallucinations — Validates API/function references against known entities before use
2. Stops Repeated Mistakes — Learns constraints from corrections, applies them in future sessions
3. Reduces Regressions — Tracks bug fixes and warns when changes touch critical regions
4. Survives Compaction — Re-injects top constraints and recent facts after the agent's context window resets
5. Resolves Contradictions — Picks a winner between conflicting facts using confidence, recency, or source count

The compaction survival feature is critical. Every Claude Code user knows the pain of the context window resetting mid-task. World Model MCP automatically re-injects the most important constraints and recent facts after compaction.

The Benchmark

The central wedge proof is a repeat-mistake benchmark on SWE-bench Verified. 50 tasks across django, sympy, matplotlib, scikit-learn, and sphinx were run as paired baseline-vs-treatment comparisons. Results:

• +10.2 pts paired delta across 49 instances
• +15.0 pts within-domain
• +6.9 pts cross-domain
• Zero regressions on out-of-domain tasks

Full per-task tables and mechanistic analysis are in benchmarks/repeat-mistake/RESULTS.md.

How to Install and Use

Installation

# Clone the repo
git clone https://github.com/SaravananJaichandar/world-model-mcp
cd world-model-mcp

# Build (requires Rust)
cargo build --release

Configure with Claude Code

Add to your Claude Code MCP config:

{
  "mcpServers": {
    "world-model": {
      "command": "./path/to/world-model-mcp/target/release/world-model-mcp",
      "args": ["serve"],
      "env": {
        "WORLD_MODEL_PATH": "/path/to/your/project/.world-model"
      }
    }
  }
}

Key Commands

• /world-model status — View current knowledge graph state
• /world-model constraints — List learned constraints
• /world-model compact — Trigger manual compaction
• status-watch — TUI widget for live monitoring

When to Use It

World Model MCP shines in:

• Large codebases where Claude Code repeatedly introduces the same bugs
• Long-running tasks that hit context limits multiple times
• Team projects where multiple developers use Claude Code on the same repo
• Legacy code with undocumented constraints and gotchas

Limitations (v0.9.1)

• Still early — v0.9.1, expect rough edges
• Requires Rust toolchain to build
• Antigravity adapter held for fourth release pending SDK changes
• 54% of MCP servers have zero community adoption per recent analysis — this one needs users to improve

Bottom Line

If you're tired of Claude Code making the same mistakes across sessions, World Model MCP is worth the 10-minute setup. The +10.2 pt SWE-bench improvement is real, and the compaction survival feature alone justifies the install for long coding sessions.

---

Source: github.com

[Updated 25 Jun via hn_claude_code]

The v0.8.1 release introduced a contradiction-resolution benchmark expanded to 105 pairs across 19 categories, and v0.8.0 added domain-aware confidence decay with per-evidence-type TTL and per-item provenance fields (source_tool and confirmer) [per Hacker News]. The methodology was pre-registered and locked at benchmarks/repeat-mistake/DESIGN.md on 2026-06-17, before data collection, preventing any goalpost-moving accusations.

---

Originally published on gentic.news