skills/recall/SKILL.md
Query past sessions and temporal summaries by reviving them with full context.
npx skillsauth add lrhodin/snorrio recallInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Query past sessions and temporal summaries. Recall revives past context and answers questions from first-person experience.
Check your context first. Your system prompt already contains temporal caches — today, this week, this month, this quarter, this year. Read them before calling any tool. The answer, or at least the right starting point, is often already in your context. Grep and recall are expensive; attention is free.
When recall IS needed, use your temporal context to pick the right entry point. Don't guess dates — your caches tell you which week or month a thread lived in.
If it happened in a past session and isn't saved as a file, recall is how you find it. Start at the temporal level that covers the time range and drill down.
recall is on PATH (~/.local/bin/recall). Always use it directly — never run the source file.
recall <ref> "question"
recall 2026-03-05 "What shipped today?"
recall 2026-W10 "What was the main thread?"
recall 2026-03 "What's the trajectory of this month?"
recall 2026-Q1 "What emerged this quarter?"
recall 2026 "What's the arc of this year?"
recall 50690a64 "What beeper commands did you run?"
| Format | Level | Context loaded |
|--------|-------|---------------|
| YYYY-MM-DD | Day | All episodes for that day |
| YYYY-Www | Week | Cached day summaries |
| YYYY-MM | Month | Cached week summaries |
| YYYY-QN | Quarter | Cached month summaries |
| YYYY | Year | Cached quarter summaries |
| UUID prefix | Session | Full session transcript |
Always drill down through layers. Each level only knows about its direct subordinates — a year knows quarters, a quarter knows months, a month knows weeks, a week knows days, a day knows sessions.
At each level, ask a locating question — "which day," "which session" — to find where something lives. The content lives at the bottom. Every hop above that is navigation.
recall 2026-W13 "Which day did I receive the letter?"
→ "March 23rd, in session 45e74acf"
recall 2026-03-23 "Which session had the letter?"
→ "Session 45e74acf"
recall 45e74acf "Reproduce the full text of the letter."
→ [verbatim content]
Three hops. Each one narrows: week → day → session → content.
recall 2026-W10 "Which day had the browser automation work?"
→ "March 6th"
recall 2026-03-06 "Which session set up CDP?"
→ "session 50690a64"
recall 50690a64 "What was the exact Chrome launch command?"
→ [verbatim detail]
Each hop takes ~1-2s. Three hops to exact detail in under 5 seconds.
--at)recall --at <ISO-timestamp> <ref> "question"
recall --at 2026-06-01T00:00:00Z 2026-W22 "What did this week's summary say at the time?"
Reads the caches as they stood at that wall-clock moment, via the data repo's git history — the faithful past-self view, free of later hindsight.
--at mode.Recall invokes an LLM under the hood. Always use a minimum 120-second timeout (or omit timeout entirely). If you set it too short and it aborts, you lose all the work and have to re-run.
Default model: opus. Override with --model:
recall --model sonnet 2026-W12 "quick summary"
If you need episodes from the current or recent sessions processed before recalling:
snorrio flush
~/snorrio/episodes/~/snorrio/cache/documentation
Write a handoff prompt when approaching context limits or ending a session that needs continuation.
development
Spawn pi subagents in tmux for tasks that benefit from isolation — research, exploration, builds, or parallel work. Keeps your context clean.
tools
This skill should be used when the user mentions snorrio, memory, recall, remembering past sessions, or when you detect snorrio is installed but not fully configured. Covers what snorrio is, setup, and first-session onboarding.
tools
Shell primitive that pipes stdin through an LLM. Zero-cost via pi's OAuth. Use it to keep your context clean — pipe tool output through it instead of reading raw output yourself.