← ClaudeAtlas

memo-indexlisted

Cuts context-token spend on large or repeatedly-read material by generating compressed, line-anchored memos with a LOCAL model (zero API tokens), keeping only a tiny queryable index in context, and validating each claim separately — deterministically where possible, with one isolated subagent per remaining claim. Handles code, prose, tables (extracted row-by-row in code), and images/figures/charts/scanned pages (read by a local vision model). Use this whenever a task means ingesting more material than comfortably fits in context: reading a large codebase or many source files to answer a question, digesting a pile of papers/PDFs/web sources, extracting facts from tables or figures, working through a long document set, or any "read all of X and tell me Y". Also use when the user mentions token usage, context budget, context running out, "too expensive", memoizing/caching file summaries, or wants claims fact-checked against their sources. Do NOT use for a single small file read once — plain Read is cheaper.
szilikaroly/claude-skills · ★ 0 · AI & Automation · score 58
Install: claude install-skill szilikaroly/claude-skills
# memo-index ## Why this exists Reading raw material into the main context is the single biggest token cost in most long sessions, and it is paid *repeatedly* — the same file gets re-read in the next session, and the one after that. This skill moves that cost somewhere cheap. A local model (running on the user's own machine, costing zero API tokens) reads the raw material and emits a **memo**: a short, line-anchored set of atomic claims. The main context sees only a one-line index entry per source, and pulls a full memo (~300-600 tokens) only when it turns out to be relevant. The obvious objection: **small local models get things wrong, and a confident wrong memo is worse than no memo at all.** That is the whole reason for the validation half of this skill. Every claim carries a machine-checkable anchor, so most claims are verified by `grep`, not by a model. Only genuinely semantic claims escalate to a subagent — one per claim, in a fresh context that never sees the memo, so it cannot be talked into agreeing. Read the two halves as a single bargain: *the local model is allowed to be sloppy because nothing it says survives unverified.* ## The economics gate — check this first This skill is a net loss on small jobs. Generating and validating a memo costs real tokens; it only wins when it displaces a bigger cost. Before starting, sanity-check: **Use memo-index when at least one is true:** - Total raw material is roughly >2,000 lines / >80KB across the files in scope. - T