MindGraph: Persistent Graph Memory for Coding Agents
TL;DR: Multiple posts discuss MindGraph, a state backend for coding agents like Claude Code and Codex, addressing the 'amnesia' and context fragmentation when sessions end, tools switch, or multi-agent concurrency occurs. The core solution persists tasks, iterations, blockers, and lessons as graph states, using symbol-based code anchors instead of line numbers to associate code context. This is crucial for long-cycle, multi-person, or multi-tool AI coding workflows.
Confirmed
- ShanRizvi introduces MindGraph for Claude Code and Codex, providing persistent graph memory across sessions.
- The system writes goals, plans, failed attempts, reasons, blockers, and lessons into a graph-managed working state; new sessions only need a short context to resume.
- Memory persistence uses code anchors: stable identities based on code symbols, not line numbers, surviving refactoring.
- The system queries a local code graph index to fetch relevant context on demand, not cramming all history into prompts.
- For multi-agent scenarios, MindGraph uses leases with epoch fencing to handle concurrency; lagging agents receive explicit expiration or conflict signals to prevent stale state overwrites.
- Performance tests show write overhead is low compared to a single LLM round trip, suitable for daily development.
Unconfirmed
- Only Claude Code and Codex are explicitly supported. A repost mentions Supermemory trying to connect Claude Code, Cursor, Opencode, Gemini, etc., but that capability is not directly confirmed for MindGraph in these posts.
Why it matters
- @Asleep-Pilot-4142 notes that switching between Cursor, Claude Code, and Codex often loses context; each tool doesn't know what the previous one did, requiring manual sync of key decisions.
- If MindGraph works, it means not just 'remembering chat history' but persisting task state, code locations, failure experiences, and concurrency control together, moving agents from one-shot tools to sustainable collaborative development executors.
2026-07-28 ~ 2026-07-29 · 8 related posts
Primary sources
- Supermemory pitches shared context transfer across Claude Code, Cursor, and Gemini — builditwithjoe · 2026-07-28
- How one developer is trying to keep Cursor, Claude Code, and Codex sharing the same context — Asleep-Pilot-4142 · 2026-07-29
- [source] MindGraph adds durable graph memory to Claude Code and Codex sessions — ShanRizvi · 2026-07-29
- [source] MindGraph gives coding agents durable memory that survives sessions — ShanRizvi · 2026-07-29
- Fixing Coding Agent Amnesia: Graph-Managed State Substrate Benchmarks — ShanRizvi · 2026-07-29
- [source] Coding Agent Memory: Using Symbol Anchors Instead of Line Numbers — ShanRizvi · 2026-07-29
- Solving Multi-Agent Concurrency: Leases with Epoch Fencing — ShanRizvi · 2026-07-29
- Graph Memory and Lease Isolation: Multi-Agent State Management in Practice — ShanRizvi · 2026-07-29