Tag
A discussion post highlighting the 'institutional memory problem' for AI agents — learned know-how scattered across vendor-controlled memory features, chat logs, and traces — referencing NeuralTrust's Oct 27 piece and InfoQ's distinction between context, memory, and knowledge, and asking how practitioners store durable, portable agent knowledge.
An analysis of 500+ AI builder community discussions finds context drift and prompt-fighting, not hallucination, are the top 2026 frustrations, motivating the launch of Skillware, a modular agent framework with deterministic execution and just-in-time skill disclosure.
The article discusses the architectural design of DeepSeek’s newly open-sourced dsh-superagent, noting that its goal of “much value and little human attention” aligns with that of the loopx project. It compares the similarities and differences between the two in state/session separation, context optimization, and router delegation mechanisms, while also promoting the related loopx RFCs.
A developer shares how 'beads,' an agent-native ticket system, solves common multi-agent workflow problems like context preservation, agent delegation, and parallel coordination by providing a shared, persistent ticket database that all agents on a network can access simultaneously.
GBrain Cloud 推出 `gbrainio` 命令行工具，可让 Claude Code、Codex 等任意编码代理一键在云端启动持久化记忆，跨多种 AI harness 共享上下文，并支持定时任务。
The authors measure how well project rules reach coding agents, finding Claude Code captures only 32% of needed rules without docs (54% with, 70% with their Segmently Modules setup), while code-indexing tools like Serena, Trail and Graphify add little. They release 'the stand', an open-source MIT-licensed context-recall benchmark kit that runs as a Claude Code plugin on real past tasks.
A founder reflects on the wave of big-tech AI agent launches in ten weeks (xAI Grok Bot, Meta Muse, OpenAI Dots, Google Gemini agent, Anthropic Claude Duet), praising their individual strengths while arguing that repeated onboarding and fragmented context is the real problem — and pitching Vestra as a shared context layer where agents plug in.
TrawMem reframes LLM agent memory systems as task-oriented role-adaptive workspaces, organizing interaction histories into threads and selecting complementary memories by their task roles, achieving consistent macro-F1 gains of 2.23–3.14 points over representative memory systems on LoCoMo and MemGallery benchmarks.
FlexA-Mem 提出一种预算自适应的智能体记忆机制，将历史一次性写入不可变存储，并为每次调用动态渲染紧凑的时间线记忆加相关细节，在 AppWorld、LongMemEval 和 LoCoMo 上显著优于 A-Mem 等现有方法。
开发者提出在项目增长过程中如何设计 monorepo 结构与开发工作流，以让 Claude Code 持续遵守架构规则并减少上下文消耗，涵盖指令管理、上下文优化与验证方法等系统性问题。
作者导出10小时Grok Bot使用数据发现89%消耗来自后台子任务、96%输入token用于重读旧上下文，并总结出5条规则，预计可节省25-30%的用量。
Ogcode 的团队分享了其用户数据洞察：开发者安装大量 skills 和 MCP 工具导致 agent 选错工具并浪费高达 77% 的 token，他们正在研发能按需动态选择工具的解决方案，并向社区征集思路。
Falconer launched to give production AI agents a unified, auto-updated knowledge base combining internal docs and public documentation, backed by GP, BoxGroup, SV Angel and angels from Anthropic, Stripe and Cursor.
Updated lecture notes for MIT OpenCourseWare 6.S950 Agency with AI, Lecture 5: The Harness, walking through a ~300-line teaching Agent to explain the principles of observability, controllability, and the question of how much Harness you actually need.
dsh-context is a DeepSeek Harness plugin that provides full context lifecycle management for agents, including a context insights page, Context/Fleet dashboards, and a /context command — enabling cross-session visualization of usage, cost, and context composition.
A free online book and GitHub repository, 'Harness Engineering 101', that teaches how to build AI agents from first principles — treating the LLM as the brain and the surrounding harness program as the body, across 16 chapters with a growing ~300-line Python toy agent.
该论文提出 RECAP 框架，通过匹配上下文干预、执行关系预测与独立策略认证，研究长程智能体何时应保留、压缩或删除历史失败记录，在 Hunyuan-A13B 上以匹配的续跑成功率将平均提示成本降低 2.13%。
The author argues that agent memory checkpoints storing a 'done' claim don't prove task completion, and asks what evidence — target-system reads, execution receipts, or task-specific checks — should be required before a long-lived agent keeps a task marked complete after a context reset.
Otaku 是一个面向创意写作与角色扮演的开源 LLM 前端工具，提供网页与终端双界面，支持本地引擎（llama.cpp、KoboldCpp、LM Studio 等）及云端提供商，并新增角色库、故事模式、图片上下文、网络搜索与上下文摘要压缩等功能。
作者分享让 AI 代理长时间运行而不丢失状态的做法：在上下文窗口写满前将工作状态保存到检查点文件，新会话读回检查点继续工作，并通过带验证的完成项来解决伪造完成的问题。