Graft

One-shot configurator that wires graft into a CLAUDE.md (global or project-local) so the agent uses the persistent memory consistently in every session. Asks the user 4 short questions (scope, caching, retrieve strategy, save strategy), then writes a `<!-- graft:start -->` … `<!-- graft:end -->` block into the chosen CLAUDE.md. Triggered by `/graft-init`, "configura graft", "set up graft for this project", "enable graft globally". Idempotent — re-running it replaces the previous block in place.

Persistent graph memory across conversations. The master skill — search BEFORE answering non-trivial questions, save AFTER solving non-obvious ones. Three companion skills handle the heavy lifting: `/memoryze` (save with smart granularity), `/recall` (escalating search query→retrieve→explore), `/memory-audit` (read-only health check). The daemon auto-starts on first command. Multi-tenant via profiles. Use this skill ALWAYS for any non-trivial technical question, framework quirk, design decision, or learned-the-hard-way fix.

Batch knowledge ingestion from external sources (a folder, a codebase, a docs tree, a set of files). Plans the ingestion as a list of well-shaped nodes with reconciled keywords, shows the user the plan for approval, then executes idempotently. Triggered by `/learn`, "ingest this folder into memory", "porta questo codice nella memoria", "acquire this documentation", "memorize this codebase". Differs from `/memoryze` (1-5 nodes from conversation) — `/learn` produces 10-200 nodes from external corpora and is plan-first by design.

Health check + maintenance audit of the graft graph. Reports hit rate, hoarding ratio, top reused nodes, never-reused nodes, stale entries, and similar-but-separate clusters that may be duplicates. Suggests concrete actions (re-save with a better title, promote to README, drop stale, narrow over-broad nodes) but does NOT modify the graph automatically — every action is proposed for the user to approve. Triggered by `/memory-audit`, "is the graph healthy", "audit memory", "check graft quality", or whenever the user wants a cleanup pass before a long-running session.

Distill the current conversation (or a specified excerpt) into N high-quality graft nodes and save them. Triggered by `/memoryze`, "save this to memory", "ricorda questo", "memorize this", or whenever the user explicitly wants the agent to commit knowledge to the persistent graph. The user can hint at granularity ("split into 3 atomic nodes", "one comprehensive node"), focus ("save the WHY, not the diff"), and target profile. Auto-classifies keywords. Prefer this over a raw `graft insert` whenever you're saving more than a one-liner.

Smart graft search that picks the right tool (query → retrieve → explore) based on the question's shape, escalates when results are weak, and presents findings ranked by confidence with the originating evidence. Triggered by `/recall`, "do we have something about X", "have we seen this before", "ricordi se", "search the graph". Prefer this over a raw `graft query` whenever the user is exploring rather than confirming a known answer.