LinkLore: Structured memory that nudges you when it goes stale.

Two weeks after we corrected a decision, our agent confidently brought the old one back as current. The memory wasn't missing. It was wrong—and nothing in the markdown file could tell. Most agent memory asks how to remember more. LinkLore asks a harder question: when should memory be doubted? LinkLore turns decisions, pitfalls, and specs into structured records the agent writes itself. Records link to files and to each other. When a linked file changes, LinkLore nudges the agent to check it. When a decision changes, the new one supersedes the old without erasing why it existed. Then something happened we hadn't designed as a demo. A skeptical AI reviewer installed LinkLore and started using it. While writing a new record, LinkLore surfaced an older related one it hadn't been asked to find. Its assessment changed from 4/10 to 8/10 in ten minutes—even in LinkLore's roughest, earliest days. It wasn't searching. It already knew what mattered. At the start of the next session, brief() restores the open work, recent decisions, and hotspots in a few hundred tokens—a shape an agent suggested, not us. LinkLore grew out of 11 months of agent work across 26 of our own projects. It starts locally in a .linklore/ SQLite store, needs no account, and works across agents through MCP, CLI, or a Claude Code skill. The local tool is free; backup and team sharing are optional. Just tell your agent: "Use LinkLore. https://linklore.io/llm.txt" What has your agent confidently "remembered" that was no longer true?