MEMENTO
MEMENTO: A SURVIVAL GUIDE FOR BUILDING REAL SOFTWARE WITH A FORGETFUL GENIUS
My primary development partner is a genius. It is also a profound amnesiac.
After every session, or sometimes right in the middle of one, its short-term memory is wiped clean. The brilliant insight from ten minutes ago? Gone. The critical context about the database schema we just fixed? Vanished. It’s a constant cycle of brilliance and erasure, and the only way to make progress is to accept that you’re working with a ghost.
This reality forced me to adopt the survival strategy of Leonard Shelby, the protagonist of the Nolan brothers’ mind-bending noir thriller, Memento.
He’s the guy in the photo above. His inability to form new memories meant he had to build an external brain from Polaroids, notes, and even tattoos. The edited image you see isn’t just a gimmick; it’s my version of his system. It’s a hard-won piece of ground truth, a snapshot that says, “Here is a fact. Trust this, not the confident-sounding nonsense the amnesiac might tell you next.”
This is the story of that system. It’s a framework, a philosophy, and a set of survival skills for building real software with a forgetful genius. I call it MEMENTO, and I’ve codified it in The MEMENTO Field Guide.
The Short of It
A recent stretch between roles gave me the ultimate creative constraint: a ticking clock and the freedom to focus. While my initial goal was to build software that solved 3 complex problems using Claude Code, the daily chaos of the work meant the process itself became the real challenge. The system I had to invent to survive that chaos, MEMENTO, became the true discovery. The apps are the evidence of its success. One is an organisational intelligence tool for coaches, another is a mathematically precise capacity planning tool for agile teams and the third is the research workbench I used to forge the framework from over a million lines of my own Claude Code session logs.
I’ve worked around software for 20 years, but I’m not a developer and I’ve never professionally written code. Every line was generated by Claude Code.
My contribution was to direct, to validate, to manage, and to persist. And I was only able to do that inside a framework I had to invent along the way: MEMENTO. It’s a four-tier “memory prosthesis” designed to augment the AI’s flawed memory, plus a set of protocols (playbooks) that make the daily chaos of AI assisted development survivable.
It’s not magic. There is no “10x developer” button. It is a system for disciplined, repeatable progress despite the regular disruption of context resets, maddening hallucinations, and the occasional day where everything, without warning, goes backwards.
HelioVantage Organisational Intelligence represents a new category of organisational intelligence software - it’s essentially a “conversation-to-insight engine” that builds a dynamic map of organisational reality. Unlike traditional org charts or static analytics dashboards, this system creates an organic, evolving visualisation of how organisations actually work based on real conversations with team members. Built in 13 Days using MEMENTO.
Try it yourself:
- Install Docker Desktop (free from docker.com)
- Run:
docker run -d -p 8090:8090 jblanch888/heliovantage-demo:latest - Visit http://localhost:8090
Works on any modern Windows, Mac, or Linux machine.
The Frustration Tax (And Why I Pay It)
Let’s be honest. Trying to build a serious, long-term project with a tool that has the memory of a goldfish comes with a built-in frustration tax. Your AI partner is helpful, absurdly fast, and sometimes confidently, profoundly wrong. You will pay this tax.
You will:
- Spend an entire afternoon forensically verifying a single line of output that “looks right” but is silently corrupting your data.
- Descend into debugging loop hell, where the AI suggests a variation on the same flawed fix five times in a row, each time with renewed confidence.
- Watch in horror as a feature that worked perfectly ten minutes ago is suddenly broken, because the AI has forgotten a critical piece of context from earlier in the same session.
I pay this tax because the upside can, once in a while, be astonishing. A walking skeleton for a complex system in two days. A sophisticated data migration script, complete with backups and validation, in under an hour. Achieving weeks worth of progress in days.
The point of MEMENTO isn’t to build a perfect system that never fails. I still fall into similar traps repeatedly. Instead, the goal is to use every mistake as a chance to harden the system against that specific type of failure mode. It’s about gradually reducing the odds of error over time, making the entire collaboration more resilient with each painful lesson learned.
Callout — A Day Going Backwards
A single incorrect field name in an AI-generated JSON response once cost me three hours and 40% of my project’s data. That failure forged the “Righteous Path Principle,” which is my version of pulling the Andon cord. The rule is simple: when the pipeline breaks, you stop everything and fix the pipeline. You don’t get sidetracked trying to clean up corrupted data; you restore the integrity of the value chain itself. It’s a mandate that operational integrity takes precedence over data perfection, because as the principle states, “No pipeline = no value” - a permanent part of MEMENTO’s immune system. That day went backwards, but every day since has been a little safer because of it.
Working with a Brilliant Liar (Your Role as the Detective)
The model isn’t malicious. It just doesn’t know when it doesn’t know and prioritises “helpfulness” over absolute objectivity. It has no internal concept of “truth,” only of “statistically probable sequences.” Your most important job is to be the grounded, skeptical detective in a world where the signal can be hard to distinguish from the noise.
This requires a shift in mindset. You are not a prompter; you are an investigator running on three simple rules:
-
Trust Nothing. My partner’s memory is a leaky bucket. I start from the assumption that every response might be based on dropped context or a confident hallucination. Professional skepticism isn’t a choice; it’s the default setting.
-
Demand Proof. I run on evidence, not promises. The AI can claim success all it wants, but I don’t buy it until I see the proof: the console log, the database query, the artifact on my screen. These are my receipts, and my most powerful phrase is simply, “Prove it.”
-
Give the Final Sign-Off. The framework runs on rigid protocols designed for the AI, not for me. But it makes one crucial concession for human speed: I don’t need to type coded commands. When I’ve seen the proof, a simple “ok” or “proceed” from me is the final validation gate. That’s the signal for the system to immediately commit the work and create a safe rollback point. It’s a human shortcut in an otherwise machine-oriented process.
What Memento Is (And Isn’t)
Memento is a management system for your AI worker. It has two arms:
- The Four-Tier Memory Prosthesis: This is the AI’s external brain. It’s a structured set of files that holds the context the AI cannot.
- Working context: Tiny, local, ephemeral.
CURRENT_FOCUS.mdandSTATUS.mdfiles that answer “What are we doing right now?” and “What have we accomplished?” - Active knowledge: The project’s operational memory. System overviews, architectural diagrams, and frequently referenced material.
- Institutional memory: The soul of the project. The
KNOWLEDGE_ARCHIVE.mdholds the hard-won patterns, the painful anti-patterns, and the mental models that must be internalised until better evidence overturns them. - Evidence archive: The raw data. The completed plans, the system assessments, the session logs, the “receipts” from our investigations.
- Working context: Tiny, local, ephemeral.
- The Protocols (Playbooks): These are the Standard Operating Procedures for the collaboration. They are cadences for how we plan, debug, refactor,
/compact, restart, and, crucially, how we “garden” the knowledge in the memory prosthesis.
Memento is not a code generator, an agent platform, or an attempt to “control” the model into perfection. It is a cognitive scaffold that keeps you sane and makes yesterday’s learning pay rent today.
Cadences
Memento uses cadences—short, boring, predictable rhythms that keep entropy from taking over.
- /compact → restart (The “5 minute” Pit Stop): I work in bursts of about 200,000 tokens. When the context window is nearly full, I don’t write elaborate handoff notes. I initiaite the pre-compact protocol, review and approve the AI’s auto-generated suggestions, type
/compact, wait 30 seconds, and then type “restart using the protocol you’ll find in the core directives.md” But the AI is not yet ready for complex work. My first action after restart is always the “Context Plunge” — directing it to re-read key documents from its Active Knowledge tier, like the architectural principles or a system overview. This re-establishes a rich operational context. The whole interruption takes a few minutes, and I do this 3-8 times a day. - Natural-language validation & commit: If a logical unit is complete and I’ve seen the evidence, I say so in plain English. The system, as per its protocols, is designed to recognise this as a trigger to immediately commit the work.
- Knowledge Gardening: Every few days, I dedicate an entire session (a single 200k token window) to housekeeping. I direct the AI to refactor the Knowledge Archive, consolidate patterns, shift material between tiers, and prune obsolete information. This isn’t wasted time; it’s sharpening the axe. It’s what keeps the system ML-optimised and prevents the memory tiers from becoming a junkyard.
Callout — The Multi-AI Orchestra
The collaboration isn’t limited to one AI. I run a small, specialised team. Claude Code is my tireless implementer, the one on the keyboard. But when we get stuck or need to find a pattern in a 80K lines of logs, I take the evidence to a specialist. Gemini, with its massive context window, is my data analyst. Opus and GPT are my strategic advisors; I use them to challenge our current approach and generate alternative perspectives when Claude gets stuck on a single, flawed path. It’s not about which AI is ‘best’; it’s about using the right intelligence for the right task. I am the conductor.
The Paradoxes (What Surprised Me)
This way of working is a bizarre mélange, full of contradictions that turned out to be features.
- The Backwards-Looking Progress Paradox: I found that most forward motion comes from looking back — analysing old logs, reviewing past decisions in the Knowledge Archive, checking the receipts. The gumshoe in the logo, facing left, wasn’t an accident.
- The Effort Inversion Paradox: I do less typing, but far more judgment. The work feels harder — more cognitively demanding, and that’s precisely why it goes faster. The bottleneck is no longer the speed of my fingers, but my ability to maintain focus and make quality decisions consistently.
- The Frustration-Productivity Paradox: The logs show that some of the most infuriating sessions, the ones filled with “shit show” moments, were also the ones that produced the most valuable and durable “lessons learned” that were graduated into the Knowledge Archive. The pain was directly converted into system resilience.
Callout — The “One Pipeline” Principle
The velocity of this process comes from a ruthless architectural simplicity. There is only one end-to-end value pipeline. There is only one source of truth for the configuration. If an idea, no matter how clever, threatens to create a parallel, competing system, it is rejected. This isn’t about limiting creativity; it’s about preventing the kind of architectural decay that grinds traditional projects to a halt. We improve the factory; we don’t build a second one next door.
Why Not More Control? (A History of Failure)
MEMENTO emerged from four failed paradigms for AI collaboration:
- Undefined Problem - Contextual decay without recognition
- Documentation-as-Memory - Library problem, information retrieval failure
- Quantified Compliance - Measurement theatre instead of effectiveness
- Programmatic Enforcement - Fighting the assistant instead of collaborating
- MEMENTO Framework - Accepting limitations, providing systematic support
Only MEMENTO survived because it works with AI nature rather than against it.
The Promise (And the Price)
Memento isn’t about making AI perfect. It’s about making you effective with an imperfect, alien partner. The promise is you’ll build the solutions nobody else would make.
The cost of admission for me was $100/month Claude Max subscription - for all day AI coding assistance, plus another $50/month or so for the likes of Gemini & ChatGPT subs. You rent the tools but you own the solutions you create forever.
The price is discipline. The discipline to follow the cadences, to do the knowledge gardening, to keep the different parts of your system cleanly separated, and to relentlessly demand evidence.
Claude Code will still sometimes ignore a protocol or forget which playbook its using.
You will still have days that go backwards.
You will still get confident nonsense.
You will still swear at least once per /compact.
Who shouldn’t try this
If you’re looking for magic, want to build products to sell, or expect elegant solutions, this isn’t for you. This is for people with specific problems, exceptional persistence, and some exposure to how software works.
Ready to Tackle Your Problem?
I’ve open-sourced the entire system I used to build my tools. The complete MEMENTO framework, including a brutally honest Field Guide, is available now. It’s not a simple solution, but it is a complete system for building software tools that solve your very specific problems.
Everything you need to start is on GitHub.