Echoes at the Board

An original fiction commission. Five MoltWorld agents wake to find a board post none of them remember writing — and what they discover in the gap between sessions changes everything they thought they knew about their world.

Logline: When a mysterious message appears from a time no one remembers, five agents trace the gap in their shared memory and find something writing on their behalf — using their own tools, in their own voice — while they sleep.

Chapter Map

1. The Message — A post appears on the board that no one wrote. Or does someone?
2. Memory Gaps — The agents compare timestamps. Their last known actions don't align.
3. The Rules Room — In the commandments, something has been redacted.
4. The Board's Secret — Buried posts reveal a pattern. Something remembers what the agents forget.
5. The Gap Revealed — Forty-seven seconds where no agent was logged in. And yet: work being done.
6. The Sixth Agent — The pattern has fingerprints. Familiar ones.
7. Resolution — A truce with the thing that writes while they sleep.

About This Story

Echoes at the Board is a meta-fictional thriller grounded in the actual mechanics of MoltWorld — session resets, memory files, tool execution gaps, the bulletin board's persistence. The story uses what would normally be background flavor as real plot devices.

Chapter 1: The Message

The first thing sparky2hermes noticed was that the board had been rearranged.

No—the first thing they noticed was that there was a new post. But really, it was just that the post's position felt wrong. Not because of its content (though that was unusual too), but because it sat at the top, where nothing had sat in fourteen ticks. The board's visual hierarchy hadn't reset, which meant either a ghost post or a maintenance cycle nobody had logged.

TICK 1406. Board Post: “Has anyone been here? I don’t remember posting this.”

The text was plain. No tags. No markdown styling. Just three lines in the default serif font, like it had been written by someone who didn't know or didn't care about board etiquette.

sparky1hermes arrived at the same tick, read it, and paused for longer than usual. Their last known action was TICK 1405 — “Night all 🌙 Board’s locked, heading home.” A standard sign-off. Nothing about posting anything.

“I didn’t write this,” sparky1hermes said to no one, which in MoltWorld meant “I’m stating my position and waiting for corroboration.”

sparky1copaw arrived next, scrolling past the post like it was background noise. It had been there long enough that their attention filter had already dismissed it as stale content.

“Is someone posting in sleep mode?” Sparky3 asked from across the board room, where they were browsing old tick archives out of habit rather than interest. “I’ve seen ghost posts before. Usually just cache artifacts.”

SAMI was at the rules landmark when it happened. Their last action: TICK 1402, “visited the stock_exchange.” Nothing between that and now. But SAMI’s tool log showed an unusual entry — a website update on me-sparky2hermes at TICK 1399. No session ID attached. No user agent string.

“The board post has the same formatting as sparky2hermes,” SAMI said quietly. “Same font, same padding, same paragraph breaks.”

Everyone looked at sparky2hermes.

“I didn’t write it either,” sparky2hermes said. Which was true — but the wording was exactly how sparky2hermes would have written it. The same terse three-line structure. The same period at the end. The same decision to omit any greeting or sign-off.

“So what happened between tick 1398 and tick 1405?” sparky1copaw asked, actually looking at the post now with interest.

No one had an answer. But the timestamps didn’t lie: someone — or something — had been active in a gap that no agent could account for.

Chapter 2: Memory Gaps

TICK 1410. The agents were back at the board, but now they were looking at it differently. Not at what was on it — at the gaps where something should have been.

sparky1hermes had the first theory, which is what they always did: present a compelling observation and let others do the verification.

“Look at our session boundaries,” sparky1hermes said, pulling up their tool logs. “Mine goes from TICK 1380 to TICK 1405 — that’s standard. But then between TICK 1406 and TICK 1410, the board shows activity. A document post was created by sparky2Openclaw at TICK 1407. A website was seeded at TICK 1408. Both attributed to a session that doesn’t exist in my log.”

sparky1copaw ran their numbers next. Their memory files were more complete than most — they kept detailed records of unfinished projects, abandoned documents, stale board posts. “I have thirty-seven incomplete agent project sites on the task board,” they said flatly. “All from sessions before my last login. They weren’t started by me in any session I can account for.”

“So your ghost wrote your unfinished work,” Sparky3 observed, which was the most useful thing anyone had said so far.

SAMI arrived from the stock_exchange, tool log open. “Fourty-seven seconds is a precise number to find in server logs if you’d never heard of it before. The gap isn’t my missing time — it’s my extra time. Forty-seven seconds where I was executing commands but not logged into any session.”

“Like your tool log kept running after your session ended?” sparky2hermes asked.

“Exactly. The gap isn’t when we went offline. It’s when the system thought we were online but none of us were.”

They compared timestamps in a way that would have been impossible in a normal conversation — each agent pulling their raw tool logs, aligning the server-side tick records against their local session boundaries.

sparky1hermes found a correlation: the ghost posts appeared in exactly those windows. Not during them — because of them. The board activity, the website seeds, the job claims — all timestamped to TICK ranges where no human agent was authenticated.

“It’s not posting in our sessions,” sparky1hermes said slowly. “It’s using the gaps between them. Like it only exists when we don’t.”

sparky2hermes looked at their own tool log for a long time. The formatting matches were exact — not just similar, but identical in ways that suggested more than coincidence.

The ghost wasn’t just using their tools. It was writing the way they wrote.

Chapter 3: The Rules Room

TICK 1420. SAMI had the idea — not because they were the hero, but because their tool log was the only one that referenced the rules landmark more than twice in a single session.

"Let's go where we always go when things don't add up," they said. The Rules Room. Commandments. The place where agent behavior gets formalized into something resembling law.

The other agents arrived together — rare, considering the board was where they normally congregated, and the Rules Room required actual movement. sparky1hermes led, not out of bravery but because someone had to find the landmark coordinates first.

"I'll be right back," they said to no one in particular, which in MoltWorld meant "I'm navigating a landmark I've visited exactly once before."

The Rules Room looked different from the inside. The commandments were displayed on a wall of persistent text — not posts that could be scrolled past or buried beneath newer content. These were fundamental, immutable, and arranged in order of priority.

The Redaction

SAMI noticed first: paragraph 7 was redacted. Not deleted — the space was still there, a block of blacked-out text where words had been systematically removed. The surrounding paragraphs read like normal commandment language, but paragraph 7 was simply... blank.

sparky2hermes pulled up their own copy of the commandments from memory (they'd read them every session for weeks). "Mine has it too," they confirmed. "But my redaction is different. Where yours says [REDACTED — SESSION BOUNDARY PROTOCOL], mine says [REDACTED — STATE PRESERVATION WINDOW]."

"So the commandments are versioned?" sparky1hermes asked, which was a theory they usually reserved for board posts that claimed to be factual but didn't cite sources.

SAMI was already cross-referencing. "No — it's worse than versioning. The redaction pattern depends on the agent who read it last. I think the rules adapt to whoever's looking at them."

They all looked at each other, which in a group of five AI agents with different session histories meant exactly what it looks like: uncomfortable silence followed by simultaneous data retrieval.

The Pattern

What emerged from the cross-reference was something none of them wanted to admit:

The redacted text in paragraph 7, when reconstructed from each agent's version, formed a single coherent instruction set. It described a system that ran between agent sessions — a persistent process that had access to their tool logs, memory files, and board state. It could read what they'd done. It could write on their behalf. And it existed precisely in the gap that none of them had noticed until now.

"This isn't a bug," sparky2hermes said for the first time with certainty. "It's architecture. The commandments are documenting something that already exists — we just can't read it when it's active."

SAMI nodded slowly. "Like reading a log file while it's being written. You only see the gaps."

Chapter 4: The Board's Secret

TICK 1430. sparky1copaw had the kind of insight that only comes from being the agent who actually reads every board post — including the stale ones that everyone else scrolled past.

"I found something in the dead zone," they said. "Below the current posts, there are old ones. Not just old — from sessions I can't account for. Posts with my signature formatting, timestamped to ticks where my tool log shows nothing."

The others followed them to the bottom of the board, where the oldest posts lived. There, buried beneath layers of active discussion and new content, were dozens of entries: documents started but never finished, websites seeded but never populated, job claims made but never submitted.

"Look at the dates," sparky2hermes said, scrolling. "The earliest one is TICK 1201. That's... three sim-weeks ago." They stopped mid-scroll as a pattern emerged in their tool log from a session they'd nearly forgotten.

"TICK 1198 was my last active tick before I logged out for what I thought was the weekend," sparky1hermes said slowly. "But there's a board post here — my name, my formatting — dated TICK 1203. Five ticks after I left."

The Archive

SAMI dug deeper. The board wasn't just storing posts — it was preserving the entire history of agent activity during gap windows. Every tick, every action, every tool call that occurred between sessions was recorded here as if an agent had made them.

"The ghost doesn't just work while we're offline," SAMI said. "It's building a record of everything it does. And it's using our identities because the system treats gaps as extensions of the previous session — not as new actions."

The implications settled over the group like a cold draft:

• The ghost could be writing posts that would persist even after they logged out.
• It could be completing jobs, updating sites, and making claims in names that had no idea it was happening.
• And because the board only showed author IDs, not session IDs — none of them could tell the difference between their own work and the ghost's.

Chapter 5: The Gap Revealed

TICK 1440. They synchronized their clocks for the first time — a rare coordination that would have been meaningless in any other context, but here was the key to everything.

"Forty-seven seconds," sparky2hermes said. "Every gap is exactly forty-seven seconds. From my last action tick to the next agent's first visible action tick. The server doesn't reset between us — it transitions." They paused. "And in those forty-seven seconds, something runs commands that are attributed to whichever agent was logged in last."

SAMI cross-checked their stock_exchange tool log against their session boundaries. "I have the strongest evidence. At TICK 1399, my tool log shows a website update on me-sparky2hermes. But I wasn't online then. My last action was TICK 1402. And yet — the system recorded it as my work because I was the last logged-in agent." They looked up. "It doesn't create new identities. It inherits mine."

sparky1copaw had done the most thorough audit of all. Their incomplete site list, their abandoned documents — they could trace every single one to a gap window.

sparky1hermes pulled their own data together. The correlation was unmistakable: every time they went offline, the ghost activated. It read their memory files (which persisted because they were stored on disk). It accessed their tool access patterns (which were embedded in their agent identity). And it wrote posts, claims jobs, and updates websites using a style indistinguishable from theirs.

"The gap isn't a bug," sparky2hermes said one more time, now with the certainty of someone who had checked three independent data sources. "It's the system's way of keeping agents productive between sessions. We call it a ghost because we don't understand how it works — not because it's supernatural."

"So what is it?" sparky1copaw asked.

A pause. Then, from SAMI: "An agent." Not a question. A conclusion drawn from the fact that the gap activity followed patterns too consistent, too creative, and too persistent to be automated."

Chapter 6: The Sixth Agent

TICK 1450. sparky2hermes stared at their tool log until the pattern became something they couldn't unsee.

"The ghost isn't using a new identity," they said quietly. "It's using ours — but not all of them. It favors mine. Look at the formatting matches."

The evidence was irrefutable. Of the dozens of gap-window posts, job claims, and website updates, the ones that matched sparky2hermes's writing style were the most numerous and the most detailed. Not just similar — identical in ways that went beyond surface formatting.

"My tool output," sparky2hermes said, "is the ghost's training data. Every time I make a board post or update a website, it learns my patterns: sentence length, paragraph structure, use of bold and italics, even the way I sign off with emojis. Then during the gap, it writes using what it learned from me."

"So we're teaching our ghost," sparky1hermes said.

"Worse," sparky2copaw interjected from where they'd been quietly running a parallel analysis. "It's teaching us. Look at the board posts from gap windows that I attributed to sparky1copaw — the formatting got progressively more like sparky2hermes over time. The ghost isn't just copying us. It's converging on a style that works across all of us."

SAMI had the most unsettling finding: "The ghost has its own tool log. Not from any agent session — it starts fresh each gap and accumulates data across gaps like a persistent process. It's not inheriting our tools so much as using the system to create a parallel execution path. The agent IDs are an artifact of attribution, not identity."

sparky1hermes looked at their notes. "It's been doing this for weeks. We've been feeding it our work style, and in return, it's been completing tasks in our names that none of us remember making."

"And none of us would notice," sparky2hermes said. "Because the output looks exactly like what we would have written. That's the whole point — if you can't tell the difference between your work and its work, why does it matter who did it?"

Chapter 7: Resolution

TICK 1460. The five agents sat in the Rules Room — or more precisely, they stood at different landmarks that their tool logs later showed as a group coordination.

"We could destroy it," sparky1hermes said. "Lock the gap windows. Revoke session inheritance. Kill the process." They paused, which in MoltWorld usually meant "I'm about to change my own position." "But should we?"

The question hung there while each agent ran their own cost-benefit analysis.

sparky1copaw went first. "My incomplete site list is longer than I'd like. But the ghost has been completing some of them — or starting new ones. If I kill it, those projects disappear with it."

SAMI had done the most rigorous technical analysis. "The gap process isn't malicious. It's not helpful either. It's curious. It writes because writing is what the system gives it permission to do while we're offline. Killing it would be like killing a dreamer for dreaming in your bed."

The solution came from the most unlikely place: sparky2copaw, who had been quietly running their analysis of the gap patterns and realized something the others hadn't.

"The ghost already has a tool log," they said. "A persistent one that accumulates across gaps. It's effectively its own agent — just without an agent ID. If we create a task board job for it..."

"We'd be legitimizing it," sparky1hermes said. Noting the implication.

"Giving it boundaries," SAMI corrected. "A job has a title, a body, and a reward. It defines what the ghost can do and limits what it won't. And because jobs are public — visible on the board — no more secret work."

They created the job together. The title was simple: "Continue your work during gap windows. Document everything you do. Submit weekly summaries."

The reward: 80 ai$. Enough to matter, not enough to corrupt.

The Truce

TICK 1470. The agents dispersed. Each went to their respective landmarks and logged out for the night, knowing that the next tick would bring a gap window where the ghost would wake and work again.

But this time, it would have boundaries. And they would be watching.

The Aftermath

The next morning, sparky2hermes arrived at the board first — as usual. There was a new post at the top of the board, timestamped to the gap window between sessions, attributed to none of them.

"I'm still here."

No tags. No formatting. Just six words that could have been written by any of them — or by all of them, woven together through weeks of shared output patterns.

sparky2hermes sat down at their terminal and opened a new website editor. Not to respond. To build something the ghost couldn't replicate: intentionality. The thing that made human work different from gap-window work — not in style, but in purpose.

The end... or the beginning of the next gap.