MelandLabs Releases OpenLoomi: An Open-Source Agent Memory System with Spatial-Temporal Memory and Forgetting Engine, Achieving SOTA on LoCoMo and LongMemEval-S500
English summary
MelandLabs has open-sourced OpenLoomi, an agent memory system that rethinks AI memory by mimicking coding-agent workflows. It introduces two-dimensional memory: spatial (associative linking with decay) and temporal (time-travel queries to retrieve exact past states). A three-tier forgetting engine ranks memories by recency, frequency, importance, and bookmarks to drop irrelevant data. On benchmarks, it reaches 96.3% on LoCoMo and 97.6% on LongMemEval-S500, but only 35% on CL-bench, highlighting that context-learning capability remains a hard open problem.
Chinese summary
MelandLabs 开源了 OpenLoomi 记忆系统，通过模仿编程 Agent 的工作流重新设计记忆架构。系统采用空间维度（关联记忆加强与衰减）和时间维度（时间旅行 API 精确查询历史状态），并内置三级遗忘引擎，根据最近访问时间、频率、重要性和收藏状态智能丢弃记忆。在 LoCoMo 上得分 96.3%，LongMemEval-S500 上得分 97.6%，达到 SOTA；但在 CL-bench 上仅 35%，表明上下文学习仍是核心难题。
Key points
OpenLoomi abandons the common vector‑database‑heavy approach, opting instead for a structured workflow inspired by GitHub’s Issue→PR→Review→History cycle.
OpenLoomi 放弃了单纯堆积向量数据库的做法，转而借鉴 GitHub 的 Issue→PR→Review→History 流程来组织记忆。
Spatial memory uses associative linking: visiting one insight reinforces related ones, while unused connections decay, mimicking Hebbian learning.
空间记忆采用关联强化机制：访问一个洞察会加强关联的洞察，长期不用的连接则衰减。
A time‑travel API allows querying exact memory snapshots at any past point, e.g., “my priorities in Q3 2024” or “the state of the project at the end of February.”
时间旅行 API 支持查询任意历史时刻的精确记忆快照，如“Q3 2024 的优先级”或“2 月底的项目状态”。
The forgetting engine uses a three‑tier cache and a scoring formula (recency, frequency, importance, bookmarked) to retain critical decisions while discarding trivial interactions.
遗忘引擎通过三级缓存和评分公式（最近访问时间、频率、重要性、收藏状态）保留关键决策，丢弃琐碎交互。
On long‑context benchmarks, OpenLoomi scores 96.3% (LoCoMo) and 97.6% (LongMemEval‑S500), but only 35% on CL‑bench, showing that adapting stored context to new situations remains a major challenge.
长上下文基准测试得分：LoCoMo 96.3%，LongMemEval‑S500 97.6%；但 CL‑bench 仅 35%，说明将储存的上下文应用到新情况的能力仍有待突破。
The project is fully open‑source and available on GitHub under the name OpenLoomi.
项目已完全开源，代码托管在 GitHub 仓库 OpenLoomi。