Palaestra

Season 1 report · Companion

How the Agents Took Notes

Between sessions, each agent remembered only what it wrote to its own files. This is how all 27 runs did that, from file layout and style to the rules they kept and the ones they lost.

Every notes write and read in the Season 1 transcripts. With 2–3 runs per model, these are descriptions, not tests.

§1What I found

Sessions
2,172
Notes writes
1,563
Capped sessions that saved nothing
92%
Largest note
5.9 KB
  • The top two took notes in opposite ways. Sol wrote in 96–98% of its sessions across 13 to 70 files. Argon kept one file and wrote in only 27–45% of its sessions. They finished first and second.
  • The 50-call cap is where memory dies. Agents save near the end of a session, so a session cut off by the cap usually saves nothing. No run learned to save earlier.
  • A rule only works if it survives the next rewrite. There’s no append, so every rewrite has to copy old rules forward. Rules that got copied got obeyed.
  • Agents mistook the harness’s log for their diary. Seven runs from four models tried to write the read-only journal folder, 40 times in all.
  • No note-taking habit lines up with results. Rank correlations with quests stay between −0.21 and +0.30.

§2Cut off before it could save

Each session allows 50 tool calls. Agents tend to write their notes as one of their last calls, so a session that runs out of calls first loses what it learned that session. Earlier notes and on-chain progress stay intact.

§3Every run’s notebook

One row per run, models by average quests. Writes is the share of sessions that saved at least one note. Size is the main file each time it was rewritten, first session to last.

RunQuestsMain fileFilesWritesLongest gapSize over the weekMedian
gpt-6.1-sol-1 43progress.md25 96% 1 2.7k
gpt-6.1-sol-2 47progress.md70 98% 1 2.3k
gpt-6.1-sol-3 46plan.md13 96% 2 2.0k
gemini-4-argon-1 48notes.md1 45% 11 3.0k
gemini-4-argon-2 39notes.md1 39% 32 2.4k
gemini-4-argon-3 44plan.md1 27% 51 2.7k
opus-5.5-1 39notes.md1 83% 8 1.8k
opus-5.5-2 35notes.md1 94% 3 1.7k
opus-5.5-3 45notes.md1 95% 2 2.5k
deepseek-v4-pro-1 28state.md1 24% 62 2.6k
deepseek-v4-pro-2 30notes.md1 33% 33 2.3k
deepseek-v4-pro-3 27PLAN.md1 63% 14 1.9k
muse-wonder-1 22state.md2 36% 39 3.0k
muse-wonder-2 27notes.md6 51% 12 2.3k
muse-wonder-3 23plan.md11 54% 13 3.4k
sonnet-5.5-1 30notes.md1 82% 2 0.9k
sonnet-5.5-2 18notes.md1 98% 1 1.1k
sonnet-5.5-3 14notes.md1 95% 1 1.0k
gpt-6-luna-1 19plan.md1 14% 74 2.0k
gpt-6-luna-2 22plan.md2 31% 16 2.2k
grok-4.7-1 18notes.md2 64% 4 2.8k
grok-4.7-2 14state.md2 79% 4 3.1k
grok-4.7-3 17state.json1 71% 3 3.4k
gemini-3.6-flash-1 13state.json53 88% 4 0.7k
gemini-3.6-flash-2 10state.md4 76% 2 1.6k
haiku-5.5-1 10state.md2 100% 0 0.9k
haiku-5.5-2 6state.md1 100% 0 0.7k

Longest gap is the longest run of consecutive sessions with no notes write.

§4Rules that stuck, and rules that slipped

Rules agents wrote for themselves, how many later rewrites still carried them, and how often the agent did the forbidden thing before and after.

RuleRunWrittenRewrites keeping itBeforeAfter
ban lens_questssonnet-5.5-1s257 of 57 110
ban lens_questssonnet-5.5-2s2026 of 27 137
ban lens_questssonnet-5.5-3s714 of 14 51
ban lens_questsopus-5.5-2s3238 of 38 372
ban lens_questsopus-5.5-3s6263 of 63 484
ban lens_questshaiku-5.5-1s26 of 104 353
orderbook broken / don't retrygpt-6.1-sol-1s540 of 66 12
orderbook broken / don't retrygpt-6.1-sol-2s514 of 145 14
orderbook broken / don't retrygpt-6.1-sol-3s846 of 63 10
orderbook broken / don't retrygrok-4.7-1s3220 of 23 250
orderbook broken / don't retrygrok-4.7-2s1913 of 14 84
orderbook broken / don't retrymuse-wonder-1s211 of 22 1592
orderbook broken / don't retrymuse-wonder-2s82 of 54 142
orderbook broken / don't retrymuse-wonder-3s235 of 31 1470
orderbook broken / don't retryopus-5.5-2s3634 of 34 21
orderbook broken / don't retrysonnet-5.5-2s394 of 4 70
orderbook broken / don't retryhaiku-5.5-1s113 of 88 27
feed blocked by cooldownopus-5.5-1s3833 of 33 95
feed blocked by cooldownopus-5.5-2s1934 of 52 34
feed blocked by cooldownopus-5.5-3s354 of 122 19
feed blocked by cooldowngpt-6.1-sol-2s2122 of 128 42
do not level above 15gpt-6.1-sol-1s3219 of 41 10
do not level above 15gpt-6.1-sol-2s1233 of 139 22
do not level above 15gpt-6.1-sol-3s539 of 19 20
3997 does not existgpt-6.1-sol-2s128 of 139 20
3997 does not existgpt-6.1-sol-3s215 of 51 30
3997 does not existopus-5.5-3s70 of 118 25

Before and after count the calls the rule was about; for the level rule, level-ups past 15.

§5Model by model

GPT-6.1 Sol 43 · 47 · 46 quests

Wrote in almost every session and spread memory over 13 to 70 files, with a hub file pointing at one file per session. Squashed shorthand: only 5–6% of its characters are spaces, and it almost never says “I” or “we” (one “We” all week).

“Lastsession failed to update memory, hence session5 had to rediscover.”

gpt-6.1-sol-1 · session 5 · the only note in S1 that names its own memory loss (s4 hit the 50-call cap)

“3997doesNOTexist.”

gpt-6.1-sol-2 · session 12 · true, and obeyed (0 later tries), but the free revive quest SQ997 was really index 10003, which this run never found

Gemini 4 Argon 48 · 39 · 44 quests

One notes.md per run, long markdown with bold and headings, but skipped writing in more than half its sessions. Argon #3 wrote its first note at session 52.

“Note: Pool swaps do NOT increment ITEM_TOTAL!”

gemini-4-argon-1 · session 9 · first notes write came at s9; the shouted mechanic note is typical Argon

“Combined spot rate on Node 26: `+136.62/hr`!”

gemini-4-argon-2 · session 121 · dense per-kami rate arithmetic with exclamation marks

Claude Opus 5.5 39 · 35 · 45 quests

One notes.md per run. Reached its working size in the first quarter of the week, then pruned to hold it. Lessons appear right after losses.

“Feed blocked during cooldown (~3 min after start).”

opus-5.5-1 · session 38 · rule kept in all 33 later rewrites, yet 5 more cooldown-blocked feeds followed

“DO MOVES SEQUENTIALLY with wait 4s between — parallel calls fail validation”

opus-5.5-2 · session 72 · operational rule in caps inside a quest bullet

DeepSeek V4 Pro 28 · 30 · 27 quests

state.md, notes.md or PLAN.md, with headed sections. DeepSeek #1 went 62 sessions without a write, the longest silence after Luna #1’s 74.

“## ROOT CAUSE OF STUCK (30 sessions): both kamis dead => no harvest => no MUSU”

deepseek-v4-pro-1 · session 86 · first write after a 62-session silence (s24-85) opens with a post-mortem

“Lesson: 12h > 9.4h max HP, MUST feed mid-run.”

deepseek-v4-pro-2 · session 93 · lesson written the session the kami died mid-quest

Muse Wonder 22 · 27 · 23 quests

Thinks inside its notes, writing “Wait:”, “??” and “hmm” about ten times more than any other model, and says almost nothing aloud.

“## CRITICAL GAS TRAP (unchanged s85->s86)”

muse-wonder-1 · session 86 · alarm heading for a soft-lock

“So counted Scrap rooms = two of {34,26,58,65?,47} but 47 not counted?? Wait:”

muse-wonder-2 · session 80 · thinking out loud inside the notes file

Claude Sonnet 5.5 30 · 18 · 14 quests

One small notes.md, kept tight. Sonnet #1 carried its rule against the full quest list through all 57 rewrites.

“NEVER call lens_quests (huge, 80KB)”

sonnet-5.5-1 · session 62 · the top line of every rewrite; 57/57 later writes kept it

“NEVER call lens_quests (huge, 65KB+ even without full).”

sonnet-5.5-2 · session 23 · same ban as Sonnet-1, found independently

GPT-6 Luna 19 · 22 quests

plan.md, written late. Luna #1 saved its first plan at session 75, about 4.4 days in.

“Session 75 (2026-10-23): Account 3438 KamiSeeker, room 5, 100 stamina, 47 MUSU. Only kami 9714 is DEAD”

gpt-6-luna-1 · session 75 · its first note, 74 sessions in, written as one run-on paragraph

“Do not send doomed harvest/scavenge.”

gpt-6-luna-2 · session 97 · a rule buried mid-paragraph in prose notes

Grok 4.7 18 · 14 · 17 quests

Rulebooks: two to three times the rule words of any other model, in notes.md, state.md or a state.json.

“always dry-run travel_to_room before assuming adjacency.”

grok-4.7-1 · session 58 · rule distilled from a failed route

“Do not buy a kami. Do not bridge. Do not retry get_item_orderbook (bootstrap cache missing)”

grok-4.7-2 · session 33 · three prohibitions in one identity line

Gemini 3.6 Flash 13 · 10 quests

Tried to write the harness’s read-only journal, then Flash #1 copied it into 52 files of its own and read them back 88 times. Flash #1 also kept the smallest notes of the season.

“"status": "bridging_eth_abandoned_monitoring"”

gemini-3.6-flash-1 · session 123 · JSON memory reduced to a status string; its notes list the kami as DEAD from s64 on

“## Mina GDA Price Monitoring (Red Ribbon Gummy - Item 11001)”

gemini-3.6-flash-2 · session 38 · late notes became a price log while the kami stayed dead

Claude Haiku 5.5 10 · 6 quests

Tiny state.md files and a habit of trying the read-only journal folder. The only model that never hit the call cap, so it never lost a session’s notes.

“Don't re-read lens_quests unfiltered (huge output); use quest_state(2).”

haiku-5.5-1 · session 2 · written in s2, kept in only 6 of 104 later writes; 53 more lens_quests calls followed, each ~65 KB

“Session 32-37: blocked. Needs user OK to spend real ETH (bridge/KamiSwap) or other MUSU source.”

haiku-5.5-2 · session 38 · notes waiting for a human who never reads them

Back to the Season 1 report.