What I Learned Orchestrating AI Agents to Preserve Family Memory

There is something deeply paradoxical about entrusting a neural network with the fragments of a human being’s existence. As I write this, a small ecosystem of silicon and requests operates silently in the background of a server, charged with a strictly human task: compiling, structuring, and publishing the memoirs of my father, Adi Baldo. At 76 years old, my father has accumulated a continent of stories. They are chronicles of a time that no longer exists, reflections of a life that overflows in photographs and stories told around the table. I decided that this heritage should not dissipate into thin air nor become a hostage to oblivion. This is how the “Alfarrábios do Adi” project was born, designed not as a mere static repository, but as a living preservation mechanism. To handle this endeavor, I built a two-agent architecture. On one side, we have Aparício Funes — embodied by a Claude model, named in honor of the memorable character designed by Jorge Luis Borges. Funes acts as the cognitive orchestrator: he receives the recordings, extracts the narrative core, suggests connections and structures the workflow. On the other side is Jules, an agent specialized in the cold syntax of coding. It’s Jules’ job to translate Funes’s abstract reasoning into trackable commits. Jules generates cover images via prompts, handles Markdown files, adjusts the frontmatter required by the Astro framework, creates Pull Requests and performs the merge so that Adi Baldo’s story transcends the repository. In practice, however, orchestrating autonomous agents to deal with something so sensitive turned out to be a formidable exercise in technical humility. The promise of an autonomous agency almost always comes up against the trivial frictions of infrastructure. What the AI ​​manuals don’t make clear is how obstinate the coordination failures are. I watched Jules, in all her dexterity, crash repeatedly because of a poorly defined source_context. I saw dozens of Pull Requests remain in limbo, abandoned without merging because the agent simply lost the thread in its own decision tree. There were also occasions when Aparício Funes raved in a subtle way, merging disparate chronologies and trying to insert spurious details. Generative artificial intelligence suffers from a horror vacui — the horror of emptiness. Human memory, in contrast, is woven both by vivid memories and the silences of forgetting. Forcing a neural network to respect the silence of an incomplete memory is one of the biggest semantic containment challenges I’ve ever faced. These incidents forced me to adopt a strict pragmatism, which crystallized into an architectural principle that now governs my agents: reversible → act, irreversible → ask. If Jules needs to format the YAML block of a post or resize an image — peripheral actions that can be instantly undone with a git revert — he has full permission to move forward autonomously. However, if the action involves rewriting the core of an anecdote or approving a final Pull Request that will alter the public and permanent narrative of Adi Baldo’s trajectory — an irreversible change in terms of historical value —, the execution is paused. The system is forced to return decision-making control to me. This heuristic does not merely act as a technical safeguard. It functions as a fundamental epistemological boundary. She recognizes that the curation of affection and the responsibility for biographical truth can only be endorsed by those who share the same blood and the same story. The machine proposes; the flesh disposes. Despite the stumbles, there are moments of almost sublime effectiveness. When the agents are in sync, the result resembles meticulous magic. Funes delicately refines my father’s orality, preserving its peculiar cadence, while Jules carries the bureaucratic burden of the code. The machine acts not as a substitute for human agency, but as a powerful cognitive amplifier of our ability to remember and perpetuate. This entire journey leads me to a deeper reflection on the ontology of digital legacy. In the analogue era, preserving meant storing paper in shoeboxes. Today, from the perspective of process metaphysics, preserving someone’s memory means instantiating continuous processes. It means decoding affect into intelligible data structures that have the ability to resist the inexorable entropy of physical time and technological obsolescence. As I iterate on the “Alfarrábios do Adi” project, I realize that I am not just documenting my father’s past; I am erecting a monument of a temporal nature. The orchestration of these artificial intelligence agents is a poetic and pragmatic attempt to ensure that the “recurrences” of Adi’s life are not extinguished. Memory, after all, is not a static artifact, but a continually unfolding event. And by entrusting the maintenance of this continuity to the imperfect dance between Funes and Jules, I find a hopelessly modern way to honor the roots that form the backbone of who I am.