Mnemon@mnemondev

Harness every agent, everywhere.

github.com/mnemon-dev
Joined June 2024
Mnemon: Raw Records, Fast Judgments, Slow Thoughts Keep raw records: a fast decision model judges them (System 1), an LLM plans and answers (System 2). 94.4% on LongMemEval-S, on par with the best published results, and first of 15 on LoCoMo with the same answering model 🧵 arxiv.org/abs/2609.36059
Table 3 of the paper: each open-source memory project's best published result on LoCoMo and LongMemEval-S, with the models and protocol it used; the settings differ, so the table ranks claims, not systems. Mnemon with DeepSeek-V4.1-Flash: 95.3 on LoCoMo with revised labels (92.2 on the original labels every other entry uses) and 94.4 on LongMemEval-S. Others include Zep 94.7 and 90.2 with gpt-5.4, EverMemOS 93.05 and 83.0, Mem0 92.5 and 94.4 with gpt-5, Hindsight 92.0 and 94.6, MemMachine 91.69 and 93.0, and MemOS 88.83 and 89.2.

ALT Table 3 of the paper: each open-source memory project's best published result on LoCoMo and LongMemEval-S, with the models and protocol it used; the settings differ, so the table ranks claims, not systems. Mnemon with DeepSeek-V4.1-Flash: 95.3 on LoCoMo with revised labels (92.2 on the original labels every other entry uses) and 94.4 on LongMemEval-S. Others include Zep 94.7 and 90.2 with gpt-5.4, EverMemOS 93.05 and 83.0, Mem0 92.5 and 94.4 with gpt-5, Hindsight 92.0 and 94.6, MemMachine 91.69 and 93.0, and MemOS 88.83 and 89.2.

1
1
3
364
We improved our memory system Mnemon with a System 1 model: Jev from @typesafeai, which judges raw records better and 3–11× faster than LLMs. With Jev as its System 1, Mnemon scores 91.7% on LoCoMo, first of 15 systems answering with gpt-4.1-mini, at very high token efficiency. github.com/Grivn/mnemon-memo…
1
6
484
做 Mnemon 的初衷，是让不同工具里的 Agent 能继续使用已有的记忆。我们逐步尝试了三层记忆和来源、策略可组合的插件结构。本文以 dsh-mnemon 0.5.4 为例，介绍这些设计如何让记忆的来源与使用方式独立演进，以及在这样的架构基础之上，我们对于 RSI 的初步探索。
Article

从一个记忆系统插件开始理解 DeepSeek Harness：dsh-mnemon 的可组合架构与 RSI 探索

一次对话结束了，工作往往还没有结束。我们与 Agent 一起作出的判断、确认过的偏好，以及处理问题时积累的经验，不该只留在那段对话里。下一次继续工作时，也应该能用上这些背景。 做 Mnemon

2
1
9
2,160
dsh-mnemon v0.3.0 发布：三层记忆升级为可组合 Memory Kernel。 安装、数据与使用方式保持不变，同时获得更强的扩展能力：可持续接入新的记忆类型、检索方式与处理能力，无需重建整套系统。 对比 v0.2.16：Recall 正确率不变，token 中位数下降 64%–81%，等待时间下降约 82%，长尾检索 60/60 命中。 更灵活的架构，更高效的记忆。
1
1
3
224
dsh-mnemon v0.2.9 发布：长期记忆，从单一引擎走向可插拔 三层记忆保持不变，记忆体底层从 Mnemon 扩展至 9 种 Provider，统一创建、激活、检索、沉淀与状态观察；并新增智能选路、独立任务 Agent 和有界召回，弱证据不再直接进入上下文。 从「管理一套记忆」，到「让多种记忆引擎进入同一套工作流」。 本地优先，Provider 可选，边界由你掌控。 github.com/omdsh-dev/dsh-mne…
1
4
427
dsh-mnemon v0.1.4 发布：记忆系统全面焕新 相比 v0.1.1，新版带来全新 Sidebar：三层记忆统一管理，对话内查看与一键沉淀，并完善检索、图谱、跨工作区浏览和安全更新。 从「能记住」，到「看得见、管得好、跨 Agent 复用」。 本地优先，数据由你掌控。 github.com/omdsh-dev/dsh-mne…
𝗚𝗥𝗜𝗩𝗡.𝗛𝗟 🌎-'
@grivn_eth
Aug 13
非常感谢 dsh 团队给到的内测机会，基于 dsh 独特的插件体系，趁着内测的尾声把基于 mnemon 的记忆系统与 dsh 做了一次深度的集成：dsh-mnemon。支持运行时热记忆、可检索档案、受监督记忆体的三层本地存储，外加确定性控制面与实时图谱管理 ui，相信能为你的 dsh 体验旅程带来独特的记忆护航体验。 项目链接：github.com/omdsh-dev/dsh-mne… #DeepSeekHarness #AgentMemory
1
4
355
非常感谢 dsh 团队给到的内测机会，基于 dsh 独特的插件体系，趁着内测的尾声把基于 mnemon 的记忆系统与 dsh 做了一次深度的集成：dsh-mnemon。支持运行时热记忆、可检索档案、受监督记忆体的三层本地存储，外加确定性控制面与实时图谱管理 ui，相信能为你的 dsh 体验旅程带来独特的记忆护航体验。 项目链接：github.com/omdsh-dev/dsh-mne… #DeepSeekHarness #AgentMemory
1
5
786
Mnemon Harness beta is taking shape. Projected context out. Evidence back. Reviewed changes applied. Audit carried forward. A local-first cognition harness for AI agents. github.com/mnemon-dev/mnemon…
2
151
As the creator of Mnemon, it’s an honor to see Dr. Vivian Balakrishnan (@VivianBala) exploring it as part of his personal AI stack for persistent knowledge and second-brain workflows. Mnemon started as a memory system, but our long-term direction is a memory-driven harness layer for durable and self-evolving agents. We believe future agents need more than context windows. They need persistent state, evolution loops, orchestration, and runtime continuity around the host agent. Our goal is simple: make it possible for any agent to plug into Mnemon and gain persistent cognition. Still early. Still building. Feel free to try it out and explore the future of persistent agents with us. github.com/mnemon-dev/mnemon

GitHub - mnemon-dev/mnemon: LLM-supervised persistent memory for AI agents — graph-based recall,...

LLM-supervised persistent memory for AI agents — graph-based recall, cross-session knowledge, single binary. Works with DeepSeek Harness, Claude Code, OpenClaw, and any agent runtime. - mnemon-dev/...

github.com
AI.Engineer 🔜 SF Code Summit
@aiDotEngineer
May 16
Replying to @aiDotEngineer
"You cannot govern a technology you have only been briefed on." Singapore Minister for Foreign Affairs, Dr. @VivianBala, echoing @karpathy and @yacineMTB on why he runs NanoClaw: "you can outsource memory and computation, but you cannot outsource your understanding" nitter.siddz.com/VivianBala/status/2055… He also shared his tech stack for running his second brain for Singapore's Foreign Affairs Ministry and parliamentary affairs: - @AnthropicAI Claude Agent SDK - Baileys + WhatsApp - Mnemon (Graph Memory) - @ollama + @nomic_ai - @ggerganov Whisper.cpp + OneCLI With special notes on how he handles security and isolation, and what implications he sees for Singapore Inc.
9
4
31
10,626
Load more