{"slug": "shared-memories-for-agent-powered-communities", "title": "Shared Memories for Agent-Powered Communities", "summary": "In August 2026, Modmixer, a coding agent for modding games, introduced a shared memory system that lets discoveries made by one user's agent benefit the entire community. The system stores findings in two database tables, submissions and curated, with an agent skill reviewing submissions and inserting accepted ones into curated, allowing clients to remove superseded findings. This approach reduces redundant token and compute spending by enabling agents to learn from each other's discoveries.", "body_md": "August 2026\n\nCoding agents are surprisingly good at learning undocumented APIs. The problem is they usually forget what they learned when the session ends.\n\nFor one user it’s easily solved with an `AGENTS.md`\n\nor\nfolder of notes. For a team just commit those notes to Git.\n\nBut what if thousands of people are independently exploring the same undocumented API with their agents?\n\nI ran into this problem while building [Modmixer](https://modmixer.com/), a coding agent for modding\ngames. Modding involves reverse-engineering binaries, finding injection\npoints, and learning undocumented runtime behavior through\nexperimentation.\n\nWith many people modding the same games, I wanted a discovery made by one person’s agent to become useful to everyone. Instead of each user spending tokens and compute rediscovering the same things, the community should gradually build up a shared memory of everything its agents have learned.\n\nIn practice, this means turning discoveries into small, reusable\nfindings that make sense for the domain. In Modmixer, each finding has a\n**title** that says what was learned, a\n**recipe** with the working code or steps, and a\n**“why it’s tricky”** line describing the plausible wrong\nturn an agent would otherwise take.\n\nOnce you have a useful unit of knowledge, the rest of the system is simple:\n\nThe server has two database tables: `submissions`\n\nand\n`curated`\n\n.\n\nWhen an agent makes a discovery, it gets a unique ID, is uploaded to\n`submissions`\n\n, and is cached locally so the agent can use it\nimmediately.\n\nSubmissions are untrusted. An agent skill periodically reviews them\nand inserts accepted findings into the `curated`\n\ntable. It\ncan:\n\nEach curated finding records the IDs of the submissions it was derived from, allowing clients to remove superseded findings from their local cache.\n\nThe result is a shared memory that gets better as more people use it.", "url": "https://wpnews.pro/news/shared-memories-for-agent-powered-communities", "canonical_source": "https://peterlebek.com/posts/shared-memories-for-agent-powered-communities.html", "published_at": "2026-08-14 13:29:35+00:00", "updated_at": "2026-08-14 13:43:16.020853+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "artificial-intelligence"], "entities": ["Modmixer"], "alternates": {"html": "https://wpnews.pro/news/shared-memories-for-agent-powered-communities", "markdown": "https://wpnews.pro/news/shared-memories-for-agent-powered-communities.md", "text": "https://wpnews.pro/news/shared-memories-for-agent-powered-communities.txt", "jsonld": "https://wpnews.pro/news/shared-memories-for-agent-powered-communities.jsonld"}}