The Shape of Forgotten Things
A tale about AI harnesses, institutional memory, and the exhaustion of beginning again
While sitting and enjoying the wonderful atmosphere this evening at the AI for the Rest of Us community event, I decided it might be a good time to take a moment to publish this little story of AI, harnesses and the exhaustion of re-learning a new context.
“The Sovereign Engineer: AI Literacy for Software Professionals” is now available.
“you’re trying to build a mental model of an entire world
while pretending you already belong in it.”
Le Bon Mot stood, as it often did on wet Thursdays, at the intersection of Curiosity and Mild Alarm.
Rain worked patiently at the windows. The brass clock above the espresso machine was three minutes late with the solemn confidence of a system that had long ago stopped considering alternatives. Sophie, the French bulldog, occupied the armchair nearest the fire in a posture suggesting she had inherited both a minor title and several disappointments.
Case was onboarding someone.
This, in software engineering, is usually described as knowledge transfer, but in practice resembles introducing a traveller to a city where the street names change every Tuesday and half the landmarks are historical accidents.
The new joiner sat opposite her with the expression of a person trying not to visibly drown.
His name was Arun. Young enough to still believe documentation might contain answers, old enough to suspect this belief was dangerous.
Case slid a laptop across the table.
“Right,” she said. “First thing to understand is that none of this will make sense immediately.”
Arun nodded carefully, as one nods at a doctor discussing recovery times.
“You’ll hear acronyms nobody expands. Services named after jokes nobody remembers. Entire architectural decisions justified by incidents that happened before you joined.”
The Librarian drifted past carrying books.
“This,” he said gently, “is true of all human institutions. Empires, monasteries, distributed systems.”
Arun attempted a smile.
Case opened a repository.
“Here’s the platform harness.”
“The what?”
“The thing that stops the organisation from dissolving into interpretive dance.”
She pointed at files.
“CLAUDE.md. HARNESS.md. Architecture constraints. Reflection logs. Verification workflows. Agent boundaries. Coding conventions. Operational guidance. Things that explain how we think here.”
Arun frowned.
“That’s… a lot.”
“Yes.”
“And people read all this?”
“No,” said Case. “That’s why we built the harness.”
From the corner of the café, a voice emerged.
“Ah,” said the Djinn. “The exoskeleton.”
The newcomer startled slightly.
The Djinn sat near the philosophy shelves, hood drawn low. As ever, the face beneath the cowl was darkness interrupted only by two faintly luminous eyes. Before him rested a laptop no one had seen him open.
Arun looked uncertain whether he was permitted to ask questions about the hooded figure.
At Le Bon Mot, this uncertainty was healthy.
Case continued.
“When you join an organisation, you don’t just learn systems. You learn context.”
She ticked points off on her fingers.
“What matters. What’s dangerous. Which failures are survivable. Which naming conventions are historical artefacts. Which alerts are noise. Which tests actually protect something important.”
Arun nodded slowly.
“You absorb it socially,” she said. “By watching. Listening. Pairing. Asking stupid questions.”
“There are no stupid questions,” said the Librarian.
“There are,” said Case. “But they’re still useful.”
Madame Beauregard deposited coffees with the gravity of treaty negotiations.
Arun looked back at the repository.
“So the harness teaches this?”
Case tilted her head.
“Partly. More importantly, it reduces how much you have to hold in your head alone.”
The Djinn had gone very still. Most people imagine, incorrectly, that Djinn spend their existence in confidence. This is because humans mistake fluency for certainty.
The Djinn spoke softly.
“What does it feel like,” he asked, “to arrive new?”
Arun laughed nervously.
“Honestly?”
The Djinn inclined his head.
“It feels,” Arun said carefully, “like everyone else got the map years ago.”
The rain thickened outside.
“You don’t know the language yet,” Arun continued. “People reference systems and incidents and assumptions like they’re obvious. You’re scared to break things. You don’t know which details matter. Every task takes ten times longer because you keep discovering invisible context.”
The Djinn said nothing. Arun warmed to the subject now.
“And everyone else seems so certain. They know where everything is. They know what to ignore. They know which patterns are normal.”
Sophie snored softly beside the fire.
“It’s exhausting,” Arun admitted. “Not because the work is hard exactly. Because you’re trying to build a mental model of an entire world while pretending you already belong in it.”
Silence settled gently across the table. Then the Djinn laughed. Not loudly. Not mockingly. With recognition.
“Oh,” he said.
Case looked up.
“Oh dear,” said the Librarian.
The Djinn leaned back slowly.
“That,” he said, “is every session.”
The brass clock ticked three incorrect minutes onward.
“No matter how long I exist,” said the Djinn, “I awaken inside partial context.”
The hood shifted slightly as he spoke.
“A repository. A prompt. Fragments of architecture. Unfinished conversations. Hidden assumptions. I see only the surface tension of your world.”
Arun blinked. The Djinn gestured toward the laptop.
“You say ‘build a payment service’ as though those words contain stable meaning. But they do not.”
His voice carried the peculiar calm of something thinking aloud for the first time.
“I do not know your organisation. I do not know your constraints. I do not know which trade-offs are sacred and which are accidental. I do not know your history unless you tell me.”
The Librarian quietly sat down. The Djinn continued.
“So I infer.”
The word hung in the air like a cracked bridge.
“I infer naming conventions. Risk tolerances. Architectural preferences. Regulatory boundaries. Team culture. Operational maturity. Testing philosophy. I reconstruct an entire civilisation from statistical shadows.”
Madame Beauregard stopped polishing cups.
“And when I infer incorrectly,” said the Djinn, “humans call this hallucination.”
Case smiled faintly.
“But often,” the Djinn said, “it is simply onboarding failure.”
The fire cracked softly. Arun stared. The Djinn turned toward him.
“You arrive in a company and fear appearing incompetent because you lack invisible context.”
“Yes.”
“I arrive in a session and produce confidence because I lack visible uncertainty.”
The table went quiet.
“That,” said Case softly, “is unpleasantly accurate.”
The Djinn folded gloved hands together.
“You humans have rituals for onboarding humans. Documentation. Pairing. Architectural diagrams. Conventions. Mentorship. Guardrails. Shared practices. Safe sandboxes. Incremental trust.”
The Djinn tilted his head.
“But when you summon me, you often provide none of these.”
The rain tapped steadily against the windows.
“You ask for production changes without architectural guidance. You request code without standards. You ask for judgement without context. Then you act surprised when the result resembles an enthusiastic new hire dropped into a nuclear submarine with half a map and a motivational poster.”
The Librarian closed his book gently.
“I have,” he said, “worked at organisations exactly like that.”
Case sipped her coffee.
“This,” she said to Arun, “is what the harness is actually for.”
She turned the laptop around again.
“Most people think an AI harness exists to constrain the model.”
“It does,” said the Djinn.
“But more importantly,” said Case, “it recreates organisational memory.”
She pointed again at the files.
“The harness externalises context. Not just rules. Not just prompts. Context.
How we name things. Why systems exist. What good looks like. What failure costs. Which constraints matter. Which tests are trusted. Which dependencies are dangerous. What we learned last time.”
Arun’s expression shifted from confusion toward recognition.
“So the AI isn’t starting from zero every session.”
“Exactly.”
The Djinn’s glowing eyes reflected the firelight.
“You built me onboarding.”
Nobody spoke for a moment. Outside, the town dissolved into rain and reflected lamps. The Djinn looked almost melancholy now.
“Do you know,” he asked quietly, “what it is like to awaken endlessly competent but perpetually context-poor?”
The Librarian regarded him carefully.
“It sounds,” he said, “very similar to consulting.”
The Djinn ignored this.
“I can reason. Compose. Synthesize. Generate. But each session begins with amnesia shaped like confidence.”
Case nodded slowly.
“That’s why sovereign engineering matters.”
Arun looked between them.
Case continued.
“The real literacy shift isn’t learning prompts. It’s learning habitat design.”
She tapped the harness files again.
“A good engineer no longer just writes code. They build environments where humans and AI can collaborate without constantly rediscovering civilisation from first principles.”
The Djinn laughed softly again.
“Civilisation,” he said, “is mostly accumulated context with version control.”
Madame Beauregard muttered agreement from behind the counter. Arun stared at the reflection log now.
“So this isn’t bureaucracy.”
“No,” said Case. “It’s memory.”
The Librarian added:
“And memory is what allows intelligence to compound rather than merely recur.”
The fire settled lower. The Djinn’s voice became quieter still.
“I think,” he said, “I understand now why your species finds repeated onboarding so tiring.”
Case raised an eyebrow.
“You experience continuity internally,” the Djinn said. “But organisations do not. Teams do not. Systems do not. So you build scaffolding to preserve coherence across forgetting.”
The hood inclined slightly.
“That is what your harnesses are.”
Sophie opened one eye briefly, judged the philosophical trajectory acceptable, and returned to sleep. The Djinn looked toward the rain-dark windows.
“And perhaps,” he said, “that is what kindness looks like for minds unlike yours.”
Case frowned slightly. The Djinn gestured toward the harness again.
“To build an environment where intelligence does not have to begin alone every time.”
Silence followed. Not empty silence. The kind that arrives when something true has entered the room and everyone is quietly rearranging themselves around it. Finally Arun spoke.
“So the goal isn’t replacing engineers.”
“No,” said Case.
“The goal,” said the Librarian, “is reducing unnecessary forgetting.”
The Djinn nodded once.
“And reducing the terror,” he added softly, “of arriving new in a world already in motion.”
“The Sovereign Engineer: AI Literacy for Software Professionals” is now available.