The deliverable
The code, the diagram, the document. Clean, versioned, shipped. They say what was done.
A My Codeuse solution
Engineering ends. Engineering memory remains.
A project produces two things: deliverables and reasoning. The first are archived with care. The second disappears with the working session that produced it.
The Engineering Memory System is a deterministic compiler. It does not search your documentation: it compiles it, verifies it, and derives from it whatever can be derived.
01 — The problem
What disappears is not information. It is the reasoning.
The consequence is not forgetting. It is reopening a debate already settled.
The reasoning
The question asked, the paths considered, the one that was ruled out and the reason for ruling it out.
Six months later
You read your own work as a stranger's. You re-reason what you had already settled, and sometimes you decide the other way.
02 — Engineering memory
Negative knowledge is worth as much as positive knowledge.
Knowing that a path leads nowhere is worth as much as knowing which one works. A path ruled out is not wasted time. Provided it survives.
Originating question. A piece of reasoning keeps the question that gave rise to it. Without it, you read a decision without knowing what it answered.
Derived relations. A relation is declared once; its inverse is computed. Writing a decision never requires updating the sources it cites.
Cited authority. A refusal names the principle it rests on. A rule that changes without the check changing is a defect, and it shows.
03 — How it works
One declared source, everything else derived.
-
01
Capture
The reasoning as it happens, not afterwards once it has already been reconstructed from memory. A decision carries its originating question, its context, and what was ruled out.
-
02
Compile
Indexes, cross-citations and states are derived. Nothing that can be computed is written by hand: a fact declared twice ends up having two versions.
-
03
Refuse
A task closed on an unmet criterion, a reference to an object that does not exist, an impossible state transition. The refusal names the rule, not a preference.
-
04
Reproduce
Same corpus, same output. No statistical model in the compilation path, no network call, no run timestamp that would move a report for no reason.
04 — Architecture
The memory lives outside the model that reads it.
An assistant can read it, cite it, propose updates to it. It does not own it, and changing model does not make it disappear.
05 — The engineer decides
The engine proposes and verifies. The human decides and commits.
An artificial intelligence can help reason, write and retrieve. It does not settle anything, and nothing enters the corpus without an engineer having wanted it there.
The value of an engineering memory lies in remaining explainable, questionable and revisable by those who come after.
06 — Definition
What it is, and what it is not.
It is
- A deterministic compiler, local and offline
- A set of commands over a local corpus
- Knowledge that is linked and verified
- A memory that outlives any AI assistant
It is not
- A RAG that would retrieve a decision without knowing it no longer applies
- A documentation platform that describes without saying why
- A decision log whose documents stay isolated
- An autonomous agent
07 — Roadmap
Where the project stands.
-
Done
The compiler: compilation, verification of rules and references (lint), index derivation (index), corpus updates by diff (patch).
-
Done
A local API relaying the compiler. The compiler proposes. The human decides. The API carries it out.
-
In progress
The console: where you read your engineering memory, where you discuss it with an assistant allowed to read it, and where you approve what will make it evolve.
-
Ahead
Several memories that talk to each other. Each corpus has its own, and a decision can cite one from another.
-
Ahead
A licence to use it. The tool has served a real project for months. Opening it up requires that it install, be understood and be repaired without the one who built it.
-
Ahead
The discipline, written down. What underpins this compiler is not a file format: it is a way of handling reasoning. A book is in preparation.
08 — Questions
What we are asked most often.
How is this different from a RAG ?
A RAG retrieves a passage that resembles your question. It does not know whether the decision it hands you is still in force, what it replaced, or who cites it. The Engineering Memory System compiles those links rather than searching for them.
Does my data leave my machine ?
No. The compiler makes no network call. The corpus, the compilation and the indexes live on the machine, and nothing is sent anywhere.
Do I need an artificial intelligence to use it ?
No. The compiler works on its own, from the command line, with no model. An assistant can read the compiled memory and propose changes, but that is an option. And in that setup, proposals go through the user's approval.
What happens if I change AI model ?
The memory does not move. It lives on your machine, in files the compiler reads; no assistant holds it. You are changing your reading tool, not your library.
Is it available ?
Not publicly yet. The compiler and its API work ; the console is being built. This page exists to establish the approach, not to sell a finished product.