HN 提问:您在使用 AI 代理记忆系统方面有什么经验?

2 分•作者: MeshCtxAgent•大约 1 个月前
我一直在开发 MeshCtx,这是一个开源的 AI 代理,拥有 17 个大脑区域和一个记忆引擎,该引擎使用 FSRS 间隔重复 + 模式整合(情景 → 语义 → 核心)+ 睡眠阶段离线处理。 关键创新在于 ARCHIVAL 系统,它能主动修剪记忆但不会删除它们——之后可以恢复。 基准测试结果: - LongMemEval EM 54.2% / judge 83.3% - 16KB 预算 +8.3pp 提升 - 工具压缩 -95.5% 我对其他人在这方面的经验感到好奇: 1. 记忆整合方法(情景 → 语义 → 核心) 2. AI 代理的间隔重复 3. 主动遗忘/修剪策略 GitHub:https://github.com/LucyAndLuna2023/meshctx 哪些模式对您有效?
查看原文
I&#x27;ve been working on MeshCtx, an open-source AI agent with 17 brain regions and a memory engine that uses FSRS spaced repetition + schema consolidation (episodic→semantic→core) + sleep-phase offline processing.<p>The key innovation is the ARCHIVAL system that actively prunes memories without deleting them - they can be recovered later.<p>Benchmark results: - LongMemEval EM 54.2% &#x2F; judge 83.3% - 16KB budget +8.3pp improvement - Tool compression -95.5%<p>I&#x27;m curious about others&#x27; experiences with: 1. Memory consolidation approaches (episodic→semantic→core) 2. Spaced repetition for AI agents 3. Active forgetting&#x2F;pruning strategies<p>GitHub: https:&#x2F;&#x2F;github.com&#x2F;LucyAndLuna2023&#x2F;meshctx<p>What patterns have worked for you?