Editing Openai/6932fed8-c6a8-8000-8477-1049d95cb85e (section)

=== Assistant: I went through your 22k+ reviews and 1.4k cards. ===
I went through your 22k+ reviews and 1.4k cards. I’ll walk you through what’s going on and then give concrete tweaks you can make.

==== 1. What’s actually giving you trouble? ====

===== 1.1. The worst “problem cards” =====

Looking at lapses (Again with previous interval > 0), total reviews, and how short the final interval still is, your biggest leeches are things like:
* Adverbs / vague expressions - 何とか / なんとか（58 reviews, 18 lapses, last interval 2d） - どうやら / どうやら（61 reviews, 16 lapses, last interval 4d） - どうも / どうも - どうにか / どうにか - 不意に / ふいに - 今更 / いまさら
* Abstract Sino-Japanese nouns / ～する nouns - 責任 / せきにん - 解放 / かいほう - 警戒 / けいかい - 真剣 / しんけん - 応援 / おうえん - 対応 / たいおう - 提案 / ていあん - 注意 / ちゅうい - 回復 / かいふく - 解決 / かいけつ - 精神 / せいしん - 現状 / げんじょう - 評価 / ひょうか - 終了 / しゅうりょう - 支配 / しはい - 印象 / いんしょう
* Look-alike / sound-alike clusters - 性格 / せいかく - 正確 / せいかく - 正解 / せいかい - 通う / かよう & 通る / とおる - 直す / なおす（vs. 直る etc.） - 防ぐ / ふせぐ - 限る / かぎる - 挟む / はさむ - 染まる / そまる

For a lot of these, you have:
* 40–70 reviews per card
* 10+ lapses
* Final intervals still only 1–6 days
* Long response times (often 10–20+ seconds on average for these specific cards)

So these really are classic SRS leeches, not just “slightly hard” items.

==== 2. What kinds of words and kanji are hardest for you? ====

===== 2.1. Script type =====

Across all reviews:
* Kanji-only words - 12,300 reviews - Again rate ≈ 25%
* Kanji + kana mixed (e.g. 解決する, 改めて) - 8,258 reviews - Again rate ≈ 21%
* Pure hiragana words (かなり, ごめん, etc.) - 1,651 reviews - Again rate ≈ 21%

Among your problem cards, kanji-only words are heavily over-represented:
* In the whole deck, kanji-only = 47% of cards.
* Among problem cards, kanji-only = 64%.

So: your main pain point is kanji compounds with no kana on the front.

===== 2.2. Kanji that appear again and again in problem cards =====

In the top ~150 problem cards, the kanji that repeat most often include:
* 解, 通, 用, 信, 意, 事
* 正, 直, 確, 態, 度, 実

These form many confusing families:
* 解放 / 解決 / 解消…
* 正確 / 正解 / 正直 / 性格 (plus all the せい〜 words)
* 通う / 通る / 交通 / 通過…

So your errors are driven by interference between very similar kanji and very similar readings/meanings — classic intermediate-advanced Japanese problem.

===== 2.3. New cards vs. difficulty =====

You also had huge spikes of new cards at the beginning:
* 2025-09-06: 165 new cards (262 total reviews, again rate ≈ 32%)
* 2025-09-07: 117 new cards
* 2025-09-08: 87 new cards
* Average over the entire period: ~18 new cards/day
* Last two weeks: ~9–10 new cards/day (much more reasonable)

There’s a clear pattern: days with lots of new cards have higher again rates and longer answer times. So overload is a big part of where these leeches came from.

===== 2.4. Time of day =====

Again rate by hour (approx):
* Best zones (fewer errors): - 08:00 → ~9% Again (few data points but very good) - 13:00–14:00 → 16–17% Again
* Hard zones (more errors): - 16:00 → 28% Again - 17:00–20:00 → 27–30% Again - 22:00 → 44% Again (few reviews but terrible performance)

You clearly remember better earlier in the day / early afternoon and are much more error-prone late afternoon and night.

==== 3. Methodological issues I see ====

===== 3.1. Overload of new material early on =====

The early days with 100+ new cards/day correlate with:
* high again rates (around 28–30% in some early weeks),
* lots of cards stuck in short-interval loops.

That’s textbook encoding overload: you tried to take in way more than your brain could consolidate, especially for abstract kanji compounds.

===== 3.2. Overall retention a bit low for FSRS =====

Across all reviews:
* Global again rate ≈ 23% → retention ≈ 77%.
* Review cards only: again ≈ 19% (retention ≈ 81%).
* Learn steps: again ≈ 47% (almost half of new-card attempts fail).

FSRS tends to work best around 85–90% target retention.
You’re running a bit “too hard mode”: okay for motivation if you accept frustration, but it helps explain the number of leeches.

===== 3.3. Only Again/Good + slow “Good” =====

You only use Again (1) and Good (3). That’s fine if:
* “Good” = quick, confident recall.

But in your data, about 12% of your Good presses take longer than 12 seconds. That usually means: your memory is weak or the answer is half-guessed. FSRS then overestimates how stable that memory is, stretches the interval too much, and you forget it next time → lapse loop.

So a lot of your “difficulties” are actually borderline recalls being graded too generously.

===== 3.4. Card design: kanji → reading only =====

All your notes seem to be:

: Front: kanji word
: Back: reading in kana

No meaning, no sentence, no hint.

This is the hardest possible type for:
* abstract kanji compounds,
* clusters like 性格 / 正確 / 正解,
* nuance-heavy adverbs (何とか, どうやら, どうも, どうにか).

For these, the brain needs context and meaning links, not just “this squiggle → this sound”.

==== 4. How to understand your learning difficulties ====

Putting it together:
# Interference between similar items - Groups like 解放 / 解決 / 回復, 性格 / 正確 / 正解, 通う / 通る, etc. are “crowding” each other. - You’re often trying to recall the exact reading of the exact compound without any strong semantic anchor.
# Abstract vocabulary without context - Words like 責任, 対応, 提案, 精神, 現状, 評価, 支配, 印象 are very context-dependent. - Without a sentence or mental scenario, they’re just floating labels.
# Overloaded initial phase - The huge burst of early new cards created a pool of “half-learned” items that became chronic leeches.
# Tired-time reviews - Many high-error reviews are in late afternoon/evening, when your attention and working memory are lower.
# Grading too leniently - Slow, hesitant recalls are marked Good, so FSRS gives intervals that are slightly too long for your actual memory strength.

==== 5. Concrete recommendations ====

===== 5.1. Handling leeches (what to do with those nightmare cards) =====

I’d treat a card as a leech if either:
* ≥ 8 lapses, or
* ≥ 40 reviews with again rate ≥ 40% and final interval still ≤ ~7 days.

You currently have ~40–50 such cards.

For those:
# Suspend from the main flow. Don’t keep hammering them in normal reviews. Tag them as leech and bury/suspend.
# Rebuild the card with more cues. For example: - Add a short Japanese sentence showing a clear context. - Add a very short meaning or keyword (even just in English or your native language). - Add a little hint about the kanji (“解 = to untie / explain”, “警 = warn”, etc.).
# Break big confusion sets apart. For things like: - 性格 vs 正確 vs 正解 - 解放 vs 解決 vs 回復 Make a mini note outside Anki where you: - write each word, an example, and a distinct explanation, - maybe create one comparison card like: > Front: 「‘solve a problem’ = ？」 > Back: 解決する > and one like: > Front: 「‘free prisoners’ = ？」 > Back: 解放する
# Re-introduce them slowly, not all at once. Once you’ve rebuilt, drip them back in at 1–2 per day max, and only when you have mental bandwidth.
# If a leech still won’t die, accept “passive only”. Maybe you treat some words as: - okay to understand in reading (you’ll get them from context), - but not worth mastering as active recall right now. For those, it’s totally fine to keep them only in immersion/reading and drop the card.

===== 5.2. Adjusting pacing and load =====
# New cards/day - The last two weeks you’re averaging ~9–10 new cards/day → this is a good range for your level. - I’d set a hard cap around 10–15 new cards/day for this kind of abstract, kanji-heavy vocab. - Definitely avoid spikes like 50–100 new in one day again.
# Daily reviews - You’re doing ~240–300 reviews/day, with peaks around 400. - If you feel mentally fine, 250–300 is okay. - On days you’re tired or busy, it’s better to reduce new cards to 0 than to skip reviews. FSRS will keep your intervals reasonable.
# Time-of-day strategy - Put your hardest reviews (mature & leeches, kanji-heavy) in a session around late morning or early afternoon (10:00–14:00). - Use late afternoon/evening sessions mostly for: - easier reviews, - “cleanup” work, - or even just reading / listening practice instead of Anki.
# Stricter Good/Again rule - Decide on a personal rule, for example: - If it takes >10 seconds, or - If you feel you “just guessed”, - or if you mentally needed to go through multiple wrong candidates first → press Again. - This will raise your short-term again rate a bit, but it will reduce long-term leeches and frustration.

===== 5.3. Specific strategies for kanji and confusing compounds =====
# Add a minimal semantic anchor to kanji-only cards. - Without bloating the card, add: - a 1–3 word gloss, or - a short phrase like “responsibility; duty” for 責任. - Even a keyword is enough to give your brain another hook.
# Use “kanji families” outside of Anki. Take your frequent leech kanji like 解, 通, 正, 確, 直, 意, etc. and: - Make a small page per kanji: - kanji, - basic meaning, - 3–4 common words you’re actually studying. - Group your problem words on those pages so your brain sees patterns instead of isolated noise.
# Targeted practice for near-synonym sets. For clusters: - 性格 / 正確 / 正解 - 解放 / 解決 / 回復 - 通う / 通る do a 5–10 minute comparison drill: - Write each word by hand. - Say the reading out loud. - Say a short example sentence (even a simple one). - Explicitly say the contrast: - “正確 = correct (no mistakes); 正解 = the right answer on a test; 性格 = personality.” You can optionally make one extra card per set that directly asks you to distinguish them (e.g. a cloze sentence where only one fits).
# Limit pure “kanji → reading” drills for the worst offenders. For leechy compounds, it might be better to flip the direction: - Back = kanji (with sentence), Front = meaning or audio or reading. - That way you test understanding & recognition rather than strict production of the reading, reducing interference pressure. You don’t have to change the whole deck; just do this for a subset of chronic leeches.
# Writing practice for a small subset of kanji. For the most problematic kanji (e.g. 解, 憶 if you have it, etc.): - Pick 5–10 kanji max. - Write them by hand from memory in a notebook a few times while saying the reading & a word they appear in. - This isn’t to become a calligrapher; it’s to deepen the orthographic and motor memory, which stabilizes recognition and reduces confusion.

===== 5.4. Working with FSRS specifically =====

Given your stats:
* Review again rate ~19% → okay but could be smoother.
* Learn again rate ~47% → new material is too hard or too dense.

I’d suggest:
# Reduce “new” difficulty rather than tweaking FSRS first. - Fewer new cards/day (which you already started doing). - Better encoding (add meaning/sentence on leech-prone vocab).
# If you can edit FSRS settings: - Make sure the target retention isn’t set ridiculously high (≥95%) for this deck. A target around 90% for vocab is usually a sweet spot: fewer lapses, less stress.
# Monitor again rate by deck after these changes. - If, after adding context + lowering new cards, your review again rate stays around 15–20% and you feel okay, you’re in a good zone. - If it’s >25% and you’re annoyed, back off further on new cards or rebuild more leeches.

==== 6. TL;DR – What to actually change this week ====

If you want a simple “do this now” list:
# Cap new cards at 10–15/day for this deck (you’re already close to this).
# Adopt a Good rule: - > 10 seconds or shaky recall = Again, not Good.
# Suspend and tag the worst leeches (those with ≥8 lapses or 40+ reviews + still short intervals).
# Rebuild 5–10 of those leeches with: - a tiny meaning, - a sentence, - and/or a note distinguishing it from its “twin” words.
# Schedule your hardest reviews in the 10:00–14:00 window when you’re clearly performing best; keep late-evening sessions lighter or skip them.
# Do one short “kanji family” session a couple of times this week: - Pick a troublesome kanji (e.g. 解 or 正), - list the confusing words, - make quick comparisons and example sentences.

If you’d like, next step I can:
* take one of your “problem families” (e.g. 解放・解決・回復 or 性格・正確・正解)
* and help you design concrete example sentences & mnemonics tailored to how you seem to be forgetting them.