recall

name	recall
description	This skill should be used when the user asks about past work or prior decisions — "what did we decide about X", "how did we do this before", "do we have notes on X", "why did we choose Y", "recall X", "check memory" — or when a task needs context from earlier sessions. Retrieves from the turing-mem store index-first, opening at most three notes.
argument-hint	[question]

tm — recall

Index-first retrieval from the turing-mem store. Fixed budget: one index read → open ≤3 notes → at most one wikilink hop → answer. Never bulk-load the store into context.

1. Guard (always first)

Run ${CLAUDE_PLUGIN_ROOT}/scripts/tm-guard.sh "$PWD" and parse its KEY=VALUE output. When this skill auto-triggered (a description match, not an explicit user request) and the guard is not fully healthy (TM_STATUS not OK, or PROJECT_EXISTS=false), skip the retrieval and at most mention /tm:init in one line — never start a repair or scaffold flow from an auto-trigger; the statuses below apply to explicit invocations.

• NEED_INIT / NEED_ADOPT: stop; follow the repair flows in <MEMORY_DIR>/references/bootstrap.md (if the store does not exist yet, the same file ships in the plugin at ${CLAUDE_PLUGIN_ROOT}/references/bootstrap.md) — or suggest /tm:init.
• NEED_MIGRATE / READ_ONLY: proceed read-only — recall never writes, so retrieval is allowed. Mention the status to the user once.
• OK with PROJECT_EXISTS=false: no project memory exists yet for this directory. Search the global scope only and say so; offer /tm:init to create the project folder.

2. Retrieval loop

1. Read the project index: <MEMORY_DIR>/<PROJECT_PATH>/INDEX.md. When the question is not project-local — user preferences, environment facts, cross-project workflow — also read <MEMORY_DIR>/global/INDEX.md.
2. Scan the ## Notes rows (newest-first, shaped - YYYY-MM-DD · <type> · [[notes/<file>]] — <gist> #tags). Match candidates on gist text, type, tags, and recency. For consolidated current state ("where were we", "current status"), prefer the scope's MEMORY.md — the index points at it.
3. Cross-project hop: when the answer likely lives in another project's memory ("how did we do X in project Y", a technique used elsewhere), read the global index's ## Projects section, pick the matching project, and read its INDEX.md. Index reads are cheap and don't count toward the note budget.
4. Open at most 3 note files total across all scopes visited (MEMORY.md counts toward the 3).
5. From an opened note, follow at most one wikilink hop, and only when the link clearly leads to the answer.
6. Answer the question, citing the store-relative paths of the notes used so the user can open them.

3. Miss fallback — tm-librarian

If the index rows (including a ## Projects hop) yield no plausible candidates, or the opened notes do not answer the question:

1. Dispatch the tm-librarian agent via the Task tool. Pass it: the question, MEMORY_DIR, and PROJECT_PATH. It greps the whole store out-of-context and returns ≤3 note paths with gists.
2. Open the returned paths (a fresh ≤3 budget) and answer as in step 2.5.
3. If the librarian also misses, tell the user nothing relevant is stored and offer /tm:remember once the answer is worked out in-session.

4. Codex fallback

No agent dispatch on Codex — run the librarian search inline, keeping output small:

grep -ril "<term>" "<MEMORY_DIR>/projects" "<MEMORY_DIR>/global" | head -20

Narrow to ≤3 files by reading only their gists (grep -h "^\*\*Gist:" <file>), then open just those. Never cat whole directories.

Rules

• Never bulk-load: no reading more than 3 notes plus 1 hop, regardless of how tempting the index looks.
• log/ checkpoints are not a retrieval surface; they are only reached through the librarian grep.
• Recall is read-only in every status — it never creates, edits, or commits store files.