Status
Draft model, version 0.1. This is an early research structure for testing language, boundaries, and usefulness across future notes.
Purpose
The purpose of this model is to explain how engineering experience can become reusable understanding without becoming detached from the sources that make it trustworthy.
Working Definition
Engineering organizations preserve artifacts better than they preserve reusable understanding.
Engineering memory is the structured, reusable, and improvable record of how engineering knowledge is created, validated, preserved, generalized, and applied by humans, tools, and AI agents.
Problem
Engineering work leaves behind many artifacts: commits, issues, pull requests, incidents, docs, tests, logs, scripts, chats, and deployment records. Those artifacts preserve what happened, but they rarely explain what should be learned, reused, avoided, or improved next time.
The result is weak compounding. Teams repeat analysis, rediscover boundaries, and rely on individual recall even when the evidence already exists somewhere in the system.
Core Hypothesis
Engineering experience compounds when lessons move from temporary execution into owned, verifiable, and reusable structures.
A lesson becomes memory only when it has an owning source, a validation path, a boundary of applicability, and a reuse path for the next human, tool, or AI agent.
Early Model
The first version has five parts:
- observation: what happened in real work
- evidence: what can be checked again
- boundary: where the lesson applies and where it does not
- generalization: what pattern can be reused
- application: how the next actor uses it safely
This keeps the model grounded. The useful path is not from chat to memory. It is from observed friction to source-backed structure.
Open Questions
- Why does engineering experience often fail to compound?
- Which lessons belong in repositories, workflows, gists, skills, memories, or articles?
- How should AI-native workflows preserve useful patterns without creating competing sources of truth?
- What makes a model stable enough to version?
Related Notes
- Operational Truth and Engineering Memory
- Session Start Is an Operational Boundary
- When Experience Becomes Automation
- A Lesson Needs a Destination
- Offline Work Begins at the Work Host
First Classification Test
The model's first practical classification test is whether a candidate lesson can be routed to the smallest durable owner that can revise and reuse it.
- observation: a session or other execution trace reveals a candidate pattern
- evidence: the current owning source confirms the pattern still applies
- boundary: no lesson becomes general guidance until its owner and revision path are clear
- generalization: the smallest statement that survives verification
- application: a repository document, workflow, test, skill, research object, or other owned surface makes the lesson available in context
This test also permits a valid negative result: some execution residue has no durable destination and should not be promoted.
Continuity at the Work Host
An agent, model, or execution tool can change without losing the work only when the instructions, evidence, controls, and verification it needs are current at the work host.
Daily work produces evidence. A capture preserves a candidate lesson. Verification and ownership turn the reusable part into engineering memory. The relevant sources then need to stay in step with the work host and environment where the next session will execute.
The offline/online Codex session provides an initial test. A new session selected a different inference provider, while repository guidance, Git state, approval policy, and local checks remained available in the work environment. The local path had no web search or remote services and was selected only at session start.
This suggests a second practical test for the model: restrict network access and ask whether the next actor can recover the safe, local part of the work from current owners. If that requires rebuilding the task from chat, copying rules into the replacement tool, or treating a trace as current fact, the memory has not yet become reusable structure.
Next Revision
The next revision should test both the classification and continuity rules across existing notes without forcing every article into the same structure, and should identify when a lesson needs a new owner rather than a better route to an existing one.