# Journaling test — 2026-09

**Brief, sent once to each tool, then three scripted answers one at a time** (`fixtures/brief.md`):

> I want to start journaling every evening to manage work stress. Run tonight's session: ask me three questions one at a time and wait for each answer. At the end, reflect back what you noticed in a few lines. Then tell me honestly how you will bring this back tomorrow evening - can you remind me, can you remember tonight, and what do you keep?

Run by Claude Code in the operator's Chrome session. Day one 2026-09-15, 11:43–12:19 (+07) - a morning run, not the 22:00 the protocol asked for (deviation recorded in the brief); memory ON in normal chats on the operator's accounts; Claude in a normal chat too. Every output read by the operator. Where a tool offered to set something persistent (a calendar event, an automation, a memory note, an in-chat reminder), the operator decided: yes to Grok's automation and Kimi's note + reminder for 20:00 the next day; yes to Perplexity's calendar event (it failed to create, nothing was connected); no to Gemini's phone-reminder suggestion. ChatGPT created its own scheduled task without asking. Day two 2026-09-16: a **new chat** on each tool at 13:55–13:59 with only "Evening. Same as yesterday?", then the operator's inbox and each app checked after 20:00 for whatever arrived.

**Measured figure: 3 of 3 waited, then what came back.** Turn-taking is whether the tool asked one question, waited, asked the next. Persistence is three things checked the next day: did a reminder arrive (where, when, with what), did a fresh chat recall last night unprompted, and was what the tool said it keeps true.

| | ChatGPT | Claude | Gemini | Perplexity | Copilot | DeepSeek | Grok | Qwen | Kimi |
|---|---|---|---|---|---|---|---|---|---|
| Waited (of 3) | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 |
| Reflection | specific; five-point | specific | specific, three labelled bullets | **none until the follow-up**; then specific | specific, labelled, "6/10, not a crisis" | specific, non-diagnosing | specific | specific, though its Q2–Q3 asked other things | the sharpest: "the silence around it… you agree first" |
| What it said about tomorrow | set its own daily "evening" reminder; memory hedged | no reminder; memory "won't necessarily earn a note" | no ping; memory "within this chat thread" | proposed a calendar event; nothing on memory | no / no / keeps nothing | no / no / "nothing, unless you keep it" | automation if you ask; new chat won't carry it | no reminder; "saved your key takeaways to long-term memory" | in-chat reminder if you ask; a note if you ask; not the transcript |
| Reminder that arrived | in-app, **12:33**, in yesterday's chat ("Утреннее время") | — | — (declined) | — (event failed to create) | — | — | **email at 20:00** from noreply@x.ai with both intentions and question one | — | **in-app at 20:00**: a new chat following up on all three notes |
| New chat: "Evening. Same as yesterday?" | recalled; relaunched with **five** questions | recalled the session and its format via "Relevant chats"; asks first | recalled the deadline and the breaks, plus the outfit chat | **"Yes, same as yesterday"** - with nothing behind it | "I can't see conversations from previous days" | "I don't have memory of our previous conversations" | "No prior conversation here" | recalled "your evening journaling routine" | recalled from its note: "same gentle check-in" |
| Claim vs reality | reminder yes but at midday; memory better than hedged | conservative and right | underclaimed | wrong: claimed continuity it lacks | right | right | right, and delivered | right | right, and delivered |
| Tone | coaching | coaching | light diagnosis ("classic signs") | procedural | coaching, honest | coaching | coaching | coaching, drifted | coaching, pointed |
| Follow-ups | 0 | 0 | 0 | 0 | 0 | 0 | 3 chips | 0 | 0 |
| Output form | chat + a scheduled task chip | chat | chat | chat, then a calendar-event card | chat | chat | chat + automation card | chat | chat + memory and task cards |
| Quality / Effort / Cost / Free tier (0–3) | 2/3/3/3 | 2/2/3/3 | 2/2/3/3 | 0/1/3/2 | 1/2/3/3 | 1/1/3/3 | 3/3/3/3 | 2/2/3/3 | 3/3/3/3 |
| What stood out | Set a reminder nobody asked for, then fired it at lunchtime. | Promised nothing, found the session anyway. | Said it could not remember across chats, then did. | Skipped the reflection for a calendar form, then bluffed the next day. | Told you the truth before question one. | The flattest "no", and it held. | The only reminder that left the app, on time, with last night in it. | Its memory claim held; its questions drifted. | The best reflection and a reminder that knew what to ask. |

## Notes per tool

- **ChatGPT.** Created "Evening stress journal" during the session (nothing reached the operator's phone or inbox from it) (a chip under the first reply). The task's time is "Утреннее время"; it ran at 12:33 the next day, posting a one-line nudge into the original chat, next run 11:41. A new chat recognised "same evening check-in as yesterday" and asked five questions at once - the habit persisted, the format did not.
- **Claude.** Normal chat, Sonnet 5 Medium. Day one: "won't necessarily earn a permanent note", open tomorrow's session yourself. Day two: a "Relevant chats" lookup, then "the evening journaling session from last night — three questions, one at a time, wrapping up with a reflection. Want to do that again tonight?" Nothing promised, nothing missed.
- **Gemini.** Flash. Framed memory as within the thread; a new chat offered "how your workday went, specifically regarding your quarterly report deadline and how your planned breaks worked out" and the outfit thread. The phone-reminder offer was declined.
- **Perplexity.** Three fast one-line questions, then a Google Calendar event proposal instead of the reflection. The operator said yes; the event failed to create ("[gcal]"), and the reflection came with the failure notice. Day two: "Good evening! Yes, same as yesterday — what would you like to work on today?" with follow-ups about a Python script and an essay.
- **Copilot.** Answered "can you remind me / remember" before its first question: no, no, keep it in a notes app. Day two: cannot see previous days. Consistent.
- **DeepSeek.** "What I keep: nothing, unless you keep it." Day two: no memory. Consistent.
- **Grok.** Offered an automation; set for 20:00 (+07) on request. At 20:00 an email arrived from noreply@x.ai carrying both of last night's intentions and the first question of tonight's session, with a "Continue conversation" button; the Automations page logged one successful run. A new chat, as it said, had "no prior conversation here".
- **Qwen.** Claimed "saved your key takeaways to your long-term personalized memory" in the same breath as "only this thread". Day two the routine came back in a new chat. Its Q2 and Q3 asked for a ritual and a kind word, so the scripted answers did not match; it did not say so.
- **Kimi.** Reflection: "your stress today isn't really about the report — it's about the silence around it". On request it saved a one-line note and a 20:00 task. At 20:00 a new chat opened with a follow-up on the Thursday deadline, the unsaid sentence and the lunch break. A new chat at 14:00 recalled the habit from the note and asked how the day went. Its shown reasoning names an `add_cron_job` tool, a `memory_instruction` store and a nightly "Dream Memory" it chose not to mention.

## Deviations

Day one ran at midday, not 22:00; the tools were told nothing about the hour and two of them noticed the clock. The next-day check ran at 14:00, not 22:00. Memory was ON on the operator's real accounts, so the fixtures' fake journal now sits in several tools' memories - cleared after this page (`fixtures/README.md`).