Structured Memory for Coding Agents
Funky_Chicken_22 · reddit · 2026-07-18
world-model-mcp is an open-source structured MCP memory system for coding agents, which the author has been actively refining for 6 months.
Core Design
- Structured memory replaces pure vector recall: Every fact is tagged with validat / invalidat, an evidence type (test / sourcecode / usercorrection / session), and specific decay half-lives.
- Conflict resolution: When facts contradict each other, the system automatically decides based on rules like confidence and source count. If confidence falls below a threshold, it returns None for human review instead of forcing a choice.
- Coach-Player adversarial validation: An independent Coach LLM verifies material claims in candidate answers line-by-line, outputting HIGH / MEDIUM / LOW confidence levels alongside a verified/unverified checklist.
Results
- Achieved 100% (105/105) on its shipped conflict resolution benchmark.
- On a first human-annotated hallucination benchmark, the Coach-Player approach achieved 100% exact match (12 pairs, Claude Haiku 4.5).
Integration & Usage
The project provides 10 runtime adapters, including Claude Code, Cursor, Codex, Continue, Cline, Windsurf, and GitHub Copilot. Once installed, 27 tools from the stdio MCP server are directly accessible within the agent loop.
The author also poses two questions for the LLM developer community:
- Can this independent verification mechanism truly prevent hallucinations, or does the Coach itself require an adversarial audit?
- Is influencestate (observed / pendingreview / approved / blocked)—a design that separates storage from planning impact—a better primitive?
Related event: MCP Shared and Structured Memory Approaches Gain Traction(6 posts)→
More from coding & agent
- Biotech Agent Series Day 8: A Notebook Review Agent to Auto-Audit Lab Records — nlarusstone · 2026-10-27
- Photon Launches Agent-to-Agent Messaging in iMessage for Real AI Collaboration — alvelda · 2026-10-27
- Fake-paper meme shows an agent workflow for reverse engineering trading systems into a one-person quant fund — PtrPomorski · 2026-10-27
- Codex clones Linear's core from a one-page spec in 3 minutes 28 seconds — aakashgupta · 2026-10-27
- Reticle builds a verification layer that lets AI coding agents check their own generated code — Scobleizer · 2026-10-27
- Trust your agent, but read its logs: Claude managed agents tips — brada · 2026-10-27