A/B on a real repo: ~93% of the premium model's codebase-reading offloaded → that model ~61% cheaper/task & ~22% faster. All-in, counting the cheap memory model: ~25% (Haiku) → ~57% (deepseek-flash), ≈full on a local model. Apache-2.0. https://t.co/zzEmwwMQes
Every Claude Code session re-reads your repo from scratch and re-pays for the same context.
I made a plugin that runs a separate, cheap model as persistent repo memory — your agent asks it instead of reloading files. 👇
#ClaudeCode#MCP
Provider-pluggable (Anthropic / any OpenAI-compatible). Zero-config on a Claude subscription — uses the subscription token on Haiku. Read-only + path-jailed. A keep-warm loop holds the prompt cache hot.
It learns passively — hooks tee the content of the files your agent reads/edits into the memory, so it warms up for free from real work and stays current. One singleton serves every session, per workspace, concurrently and over time.
@Plinz Best analysis I've read on the concept of conspiracy theory. My 2c: I would prefer a view whereby everything is a theory, and what distinguishes them is the degree that it is grounded of truths/axioms, either directly or via some rational/mathematical derivation process.
@Plinz I don't believe there is such a thing as conspiracy theory. They are just theories. Every theory has a probability to be true, partially or in whole, and no-one really knows for sure.
@dr_eeks The vaccine provides protection, but at the same time Might harm the body. It would be interesting to see the general mortality between those three clusters.
@MrsSarahMurray @carlheneghan Not much. Need to look at lots and lots of graphs and articles and research papers just to reach 60% confidence level. But that's already better than pure chance. Mainstream narrative is simply a way to short-circuit into a conclusion based on trust to authority, which saves time
@xstefanou Προτείνω στους αναγνώστες να διαβάσουν τα πρακτικά αυτής της δίκης ICAN Vs CDC, για να έχουν γνώση της "άλλης πλευράς" https://t.co/c3Kj26tH0J