Menu
Recall past work across Claude Code and Codex sessions. Finds decisions, research, fixes, and conversations by intent.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Orient on a project or project family. Shows recent activity, milestones, and commits from event logs.
Open the Session Cartographer Explorer web UI for visual browsing of session history.
Root-cause-first diagnosis gate. Forces a written hypothesis with evidence before any bug-fix code change, and logs it to the event log for later recall.
Strategic session-end preservation. Captures decisions, discoveries, and state that the automatic hooks miss.
| name | remember |
| description | Recall past work across Claude Code and Codex sessions. Finds decisions, research, fixes, and conversations by intent. |
| allowed-tools | ["Bash","Read"] |
Recall past work from the shared Claude Code and Codex session history. The consumer and producer do not need to match: Claude can recover Codex work and Codex can recover Claude work. The user is trying to recover context — a decision, a fix, a paper, an approach — not run a database query.
Before running commands, resolve ROOT to the Session Cartographer plugin
root. Prefer CARTOGRAPHER_ROOT, then CLAUDE_PLUGIN_ROOT or PLUGIN_ROOT.
If none is set, derive it from this skill's reported base directory (../..
from skills/remember). Use the conventional checkout only as a legacy
fallback. Verify that $ROOT/scripts/cartographer-search.sh exists.
Hooks determine what's in the searchable index:
CARTOGRAPHER_LOG_TOOL_USE=true)--transcript. Expensive (per-query awk over 100MB+ files); use only when the event logs + semantic index both missResults include faceted summaries (project, source, event type, time range). Use these to narrow searches:
--project <name> — filter to a specific project--since WHEN / --before WHEN — temporal filter. WHEN accepts natural-language phrases (today, yesterday, "this morning", "this afternoon", "this evening", tonight, "this week", "last week", "this month", "last month"), relative durations (7d, 2h, 30m, 1w, 3mo), or absolute dates (2026-04-20). When the user mentions time — "what did I work on Wednesday", "this morning's debugging session", "last week's audio fixes" — translate to the matching --since flag rather than searching unbounded.feature, fix, refactor, enhancement, docs, test, chore, perf, ci, style, revert, or other. These appear in summaries as [feature] Commit abc1234: ... and are searchable as keywords.git_commit, research_fetch, research_search, milestone_session_end_*, milestone_compaction_*, tool_file_edit, tool_bash| User says | Use |
|---|---|
| "today", "earlier today", "this morning's work" | --since today |
| "yesterday", "last night" | --since yesterday |
| "this afternoon" | --since "this afternoon" |
| "this week", "the past few days" | --since "this week" |
| "last week" | --since "last week" |
| "this month" | --since "this month" |
| "last month" | --since "last month" |
| "the last hour" | --since 1h |
| "the last few hours" | --since 4h |
| specific date like "April 20th" | --since 2026-04-20 |
| "between A and B" | --since A --before B |
Beyond ~30 days, transcript files are deleted by Claude Codes default TTL — event-log results still surface but the "read the transcript" step (Step 3 below) will hit a missing file. When that happens, present the event metadata as the answer and note that full context isn't available.
--threadHooks link successive events from the same session into a parent_event_id chain (the prior event must be in the same session and within 60s). Over time this turns the JSONL from a flat log into a graph you can traverse.
When the user asks "show me how I got to X", "what led up to that commit", or "what was the chain of work around Y", run a normal search to find the anchor event_id, then walk the arc:
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" _ --thread evt-xxxxxxxxxxxx
The first argument is ignored when --thread is set (pass any placeholder like _). Output is the full ancestor + descendant arc sorted by timestamp, with the supplied event marked ★. Present it as a coherent timeline rather than a search result.
Procedural questions — "how do I deploy X", "what's my release process", "the cloudflare config dance" — aren't similarity matches, they're maneuvers. The co-occurrence graph (scripts/cooccurrence-graph.js) detects recurring technical maneuvers and which projects run them. Two entry points, both alias/partial-tolerant:
# Query names a project → which maneuvers it runs + which projects share them
node "$ROOT/scripts/cooccurrence-graph.js" --maneuvers <project>
# Query names the procedure → which projects run it (e.g. cloudflare, release, merge, overleaf)
node "$ROOT/scripts/cooccurrence-graph.js" --signal <maneuver>
The map is an index, not a command store — it deliberately holds no commands and no secrets. Recovering the actual invocation is the same "map → territory" move as reading a transcript in Step 3 (so it's exempt from the no-freestyle rule below): grep the changelog for the matching project + maneuver marker.
jq -r 'select(.project=="<project>" and (.summary|test("wrangler pages deploy|gh release create|netlify"))) | .summary' ~/Documents/dev/changelog.jsonl | sed 's/^Ran: //' | tail -3
Present the recovered command(s) with project + recency. They're the user's own past invocations — don't echo any embedded tokens / account IDs gratuitously.
When a recall centers on one project but the work spans a thread, surface the co-active siblings so you can widen the search. The graph knows, e.g., approach-retreat co-threads with allserp-paper / ettac-paper:
node "$ROOT/scripts/cooccurrence-graph.js" --related <project>
Use it when results cluster on one project and the question is open-ended ("what was I doing around the AOI work") — offer to pull the related threads into the search.
Hooks emit a salience score per event ([0..1]). /wrapup milestones (0.9), feature/fix commits (0.7), and research-paper fetches (0.7) outrank routine bash commands (0.2) and chore commits (0.4). Salience multiplies into the RRF score, so deliberate strategic moments naturally rise to the top of results without any extra flag.
Write-time salience is a static prior; the access ledger makes it a learned posterior. When you actually use a result — read its transcript in Step 3 — record the access with --touch. Reuse refreshes the event's recency and compounds a frequency boost (capped at 2×), so events that keep proving useful rise across future sessions. Results that have been reused before show a (used xN) tag — that is why they may rank above fresher events. Serving alone records nothing: searching is free, using is vouching.
When you call /remember repeatedly in the same session, the script automatically suppresses event_ids that were returned in earlier calls — so each subsequent call surfaces fresh material rather than re-returning the same top-K. Activated whenever CARTOGRAPHER_SESSION_ID or the backwards-compatible CLAUDE_SESSION_ID is available in skill context.
If you actually need to re-cite an event from a prior call (the user is asking about something you already showed them), pass --all to bypass suppression for that single call:
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" "<terms>" --all
To wipe the per-session served list entirely (rare; only when starting a genuinely fresh investigation): pass --reset-served.
Do NOT freestyle grep or jq commands. Always use the unified search script.
Think about what the user is trying to recall, then translate to search terms.
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" "<search terms>"
If the user mentioned a specific project, add --project <name>:
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" "<terms>" --project scrutinizer
For more results, add --limit 25 or --limit 50. If the user says "more" or "keep going" after seeing results, re-run with a higher limit:
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" "<same terms>" --limit 30
Wildcard prefix search works: shader* matches shader, shaders, shaderlab, etc.
If event logs + semantic come up empty and you genuinely need raw transcript keyword matching, opt in with --transcript. Expect it to be slow; some sessions are 100MB+.
Show results as-is from the script output. Keep it scannable.
Search results are summaries. When you need full context, read the transcript file directly — don't wait for the user to ask.
Find the transcript path — it's in the search result as transcript:. Prefer
that path because it is provider-neutral. If missing, resolve from session:
across both provider stores:
find ~/.claude/projects -name "<session-id>.jsonl" 2>/dev/null
find ~/.codex/sessions -name "*<session-id>*.jsonl" 2>/dev/null
For a Claude transcript, read around the relevant moment:
jq -c 'select(.type == "user" or .type == "assistant") | select(.message.content | type == "string") | {type, timestamp, content: .message.content[:500]}' <transcript_path> | grep -A5 -B5 "<keyword>"
For a Codex transcript, retained user and assistant messages are currently in
event_msg and response_item records:
jq -c 'select((.type == "event_msg" and .payload.type == "user_message") or (.type == "response_item" and .payload.type == "message")) | {type, timestamp, payload}' <transcript_path> | grep -A5 -B5 "<keyword>"
Or jump to a specific message by UUID:
jq 'select(.uuid == "<uuid>" or .parentUuid == "<uuid>")' <transcript_path>
The search result is the map. The transcript is the territory.
When a result summary, transcript, or --thread arc actually contributes to the answer, touch the event_ids whose context you used — this is both the promote-on-reuse moment and the explicit result-use signal:
CARTOGRAPHER_PURPOSE=remember bash "$ROOT/scripts/cartographer-search.sh" _ --touch <event_id>[,<event_id>...]
Touch only what you used, not everything that was served. A result you read and discarded as irrelevant should NOT be touched — false vouching pollutes future rankings.
/remember that paper about foveated rendering
/remember what we decided about the shader approach
/remember the commit that fixed blur
/remember Blauch collaboration notes
/remember recent feature commits --project scrutinizer
/remember what was I working on last session
/remember what did I do this morning on Psychodeli
→ bash cartographer-search.sh "Psychodeli" --since today
/remember the audio reactivity work from last week
→ bash cartographer-search.sh "audio reactivity" --since "last week"
/remember Wednesday's debugging session
→ bash cartographer-search.sh "debug" --since "last week" --before yesterday
/remember show me how I got to that fix
→ bash cartographer-search.sh "the fix" --limit 5 (find anchor event_id)
→ bash cartographer-search.sh _ --thread evt-xxxxxxxxxxxx
/remember how do I deploy movies-mindbendingpixels
→ node cooccurrence-graph.js --maneuvers movies-mindbendingpixels (→ cloudflare-pages)
→ recover the invocation from the changelog, present it
/remember my cloudflare deploy process
→ node cooccurrence-graph.js --signal cloudflare
/remember what else was I working on around the approach-retreat AOI work
→ node cooccurrence-graph.js --related approach-retreat (→ allserp-paper, ettac-paper…)