MemoraXCode Enables Long-term Memory and Skill Reuse for Coding Agents
机器之心 · wechat · 2026-08-18
Despite the growing capabilities of Coding Agents, they struggle with long-term development: unable to retain project experience across tasks or platforms (e.g., switching from Codex to ClaudeCode), leading to repetitive exploration. MemoraXCode addresses this by building a local codebase memory and cloud long-term memory system.
Key Features:
- Smart Memory Management: Goes beyond saving history to learning "what is worth remembering" via training, automatically recalling relevant info in new chats to prevent critical experience loss during compaction.
- Procedure Memory Skills: Automatically extracts reusable solution processes from historical development traces, forming structured engineering experiences (Skills).
- Benchmarks & Results: Ranked 1st on the AML Coding Track with a score of 62 (10% improvement over Claude Mem). In a 3-hour complex task, Procedure Memory boosted scores from 11.71 to 70.30 and reduced API costs by 22.6%.
Currently supports Codex, ClaudeCode, DeepseekHarness, and OpenCode, evolving memory capabilities through data-driven training rather than rigid rules.
More from coding & agent
- One night of vibe coding: open-source Reachy robot runs 100% local with Qwen brain — Scobleizer · 2026-08-18
- Azure Cosmos DB MCP Toolkit Enables Natural Language Agents — usamawahabkhan · 2026-08-18
- Local Reasoning Configs: Managing Qwen's Thought Budget — Altruistic_Heat_9531 · 2026-08-18
- Tencent Hunyuan's UI-Mate: Open-Weight GUI Agent Hits SOTA on Computer-Use Benchmarks — Tencent-Hunyuan · 2026-08-18
- 97-page survey by LSTM father maps self-improving agents: memory evolution and tool creation — 大模型之路 · 2026-08-18
- Using Grok to check GitHub Org name availability: a time saver — Daniel_Farinax · 2026-08-18