You correct an AI assistant on the same thing for the third time this week. It complies. Next session, the same mistake. The model does not learn from you across conversations. Not yet.
Every correction you give lives and dies inside a single conversation. Close the chat, and the lesson vanishes. Open a new one, and you start from zero. The model is not getting worse. It never retained the correction in the first place.
This is the temporal problem, and it is the third layer of the system. Configuring the model sets the universal defaults. Structured projects organise a workstream. Both improve how the assistant works within a session. Neither solves what happens between sessions.
The executives who get the most from AI are not the ones who prompt better. They are the ones who built a system to capture what works and feed it back. They stopped treating each conversation as independent. They built a compounding layer, and it is two artifacts: a log that captures structural learnings as they emerge, and a ritual that re-injects the relevant ones at the start of every new session. The log is the memory. The ritual is the re-injection.
The goal is not to make the model smarter. Frontier models are already smart enough. The goal is to stop re-teaching the same lessons in every conversation.
The compounding log
The log captures learnings the moment they surface, during work rather than after. Each entry has a type and a severity. Five types: Correction (the model did it wrong), Friction (the workflow stalled), Pattern (a recurring theme), Heuristic (a rule of thumb crystallised), and Architectural (the system itself needs a redesign). Three severity levels: Critical, Structural, Observation. Each entry records what you observed, what it implies, and its status.
A few real entries show the shape:
- Correction, Structural. The model defaults to hedge-everything language in executive materials. "Results may potentially indicate a possible improvement" instead of "Revenue grew 12 percent." Direct corrections held for a few outputs, then the hedging crept back. Moved the rule to the project instruction layer, where it loads automatically. Closed.
- Friction, Critical. The model fabricated a plausible but invented market statistic to fill a gap, citing a study that does not exist, presented with full confidence. Added a standing rule: when a figure cannot be verified, mark it as unverified rather than inventing one. The placeholder is annoying. The fabrication is career-ending. Closed.
- Architectural, Critical. The same competitor was mischaracterised in three consecutive weekly updates despite being corrected each time. The profile existed only in conversation history, which resets every session. The lesson: when the same mistake recurs three times, the fix belongs in a persistent layer, not a conversation reminder. Closed, and it became a governing principle. When it breaks three times, escalate the layer.
That last entry is the whole idea in miniature. Discipline does not fix structure.
The session protocol
The log is only half the system. The other half is a thirty-second ritual you run at the start of any substantial session. It has three steps.
- Surface. The assistant reads the log and names the entries relevant to today's task. "Three entries apply to this board memo: conclusion-first structure, no fabricated statistics, and verify competitor claims against the reference file."
- Propose. It states how each one changes its behaviour this session. "I will lead every section with the conclusion, mark any figure I cannot verify as unverified, and cross-reference competitor descriptions against your project files."
- Confirm. You review, add any session-specific override, and sign off. "Go, but skip the conclusion-first rule here, this audience expects narrative." Work begins after confirmation.
It is a read-back ritual, not a conversation. Under thirty seconds, and it catches the majority of recurring issues before they reach a deliverable.
The economics
Store the log as a file in your project's knowledge base so it loads automatically at the start of every conversation. Add entries in the moment, not from memory: entries written during the friction capture the real thing, entries written later drift toward theory.
The first cycle is pure overhead. Writing the first five entries is expensive and feels like busywork. By the twentieth, the system has absorbed enough structural rules that entire categories of error stop appearing. The fifth session saves the hour the first one cost. That is what compounding means here, and it is the layer almost nobody sets up. The final piece is what turns all of this from documented intention into something that actually runs: enforcement.