{"slug": "project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents", "title": "Project memory is not chat history: a tiny handoff layer for AI agents", "summary": "A developer built Resume the Scene, a small, model-agnostic project-memory layer that lets AI agents hand off work between sessions, machines, or workers without replaying prior chat history. The MIT-licensed tool stores task state in .rsm frames alongside the repository, enabling another worker to deterministically recover what to load, what to skip, and what to do next, with a handoff example executed in CI on Python 3.11, 3.12, and 3.13. The developer is seeking feedback on what state would be needed to trust a different agent to continue a real project.", "body_md": "I kept running into the same failure mode when switching between AI workers, browser sessions, or machines: the files were still there, but the exact **working point** was gone.\n\nConversation memory can tell a model what was said. RAG can retrieve relevant documents. Neither one is a durable answer to four operational questions:\n\nI built **Resume the Scene** as a very small, model-agnostic project-memory layer for that gap.\n\nGitHub: [https://github.com/louisen0o0/resume-the-scene](https://github.com/louisen0o0/resume-the-scene)\n\nThe project keeps the handoff in files that live with the repository. A task state can look like this:\n\n```\n@task{id:#t42|goal:#g1|state:active|done:[#e69]|next:#a7}\n@msg{op:resume|task:#t42|load:[#current,#e81]|skip:[#e69]|next:#a7}\n```\n\nThe important part is not the syntax. The important part is that another worker can deterministically recover **what to load, what to skip, and what to do next** without replaying the prior chat.\n\nResume the Scene is intentionally thin:\n\n`.rsm` frames;\nThere is no vector database requirement, no model-specific identity in the core protocol, and no requirement to archive prompts or responses.\n\nThe repository now includes a handoff example where one worker has already completed evidence `#e2`, and the next worker resumes from the same task state:\n\n```\nresume-scene checkpoint examples/handoff\nresume-scene resume examples/handoff\n```\n\nThe resume result tells the next worker to skip completed evidence and continue with the next action. That fixture is executed in CI on Python 3.11, 3.12, and 3.13.\n\nThe implementation is small enough to change. The state model is the part I want challenged before it grows.\n\nIf you regularly hand work between coding agents or long-lived AI sessions, I would especially like feedback on this question:\n\n**What state would you need to trust before allowing a different agent to continue a real project without replaying the previous conversation?**\n\nThe project is MIT licensed and the roadmap is public in GitHub Issues.", "url": "https://wpnews.pro/news/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents", "canonical_source": "https://dev.to/louisen0o0/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents-2n03", "published_at": "2026-09-17 16:05:02+00:00", "updated_at": "2026-09-17 16:23:07.656262+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "agent-protocols"], "entities": ["Resume the Scene", "GitHub", "Python"], "alternates": {"html": "https://wpnews.pro/news/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents", "markdown": "https://wpnews.pro/news/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents.md", "text": "https://wpnews.pro/news/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents.txt", "jsonld": "https://wpnews.pro/news/project-memory-is-not-chat-history-a-tiny-handoff-layer-for-ai-agents.jsonld"}}