People who are close to me know very well that I quickly get frustrated if have to repeat myself. I got much better through out the years but the internal feeling never went away. Imagine how that plays with LLMs 😅. I never liked the "oh context got full, let's summarize" approach as it was very lossy. Many people swear by staying in the context with short sessions to get great results from LLMs. Imagine my excitement when I saw @mastra announcing their new memory approach: https://mastra.ai/blog/use-rag-f... Here, my main takeaway was how important was temporal clues and some search tool for past chats (RAG or not). Then they followed up with something even simpler and powerful: https://mastra.ai/blog/observati.... Around the same time folks from @sanity_io posted about their "distillation" approach, similar to Mastra: https://www.sanity.io/blog/how-w... I've been itching to try these in my @opencode setup and 2 weeks ago, with @sentry's AI Hackathon, I pulled the trigger on my OpenCode memory plugin: OpenCode Lore. I definitely messed up my sessions more times than I want to count and burnt some tokens when I broke caching for a few days. But the past week has been amazing: no compactions, no explanations, no skills. We just talk on and on and on and it just works. Keep in mind that this is still very experimental but if you dare to try, please do share your feedback!

OpenCode Lore

Stop re-explaining your codebase!

Stop re-explaining your codebase!