When the Reasoning Doesn't Survive

Source: Dev.to

Capturing Session Reasoning Before It Evaporates

I built a skill called /mark to write a session anchor—a lightweight markdown file that records a deliberative moment: what the reasoning was, why that framing was chosen, and what remained uncertain. The goal is to capture insight at the moment it occurs, rather than reconstruct it later from notes.

First Test

The first anchor I wrote was for an epistemic research framing developed earlier in the same session. By the time I ran /mark, the session had already undergone one compaction pass. Consequently, the reasoning I was trying to capture came from the compaction summary, not from the live exchange.

• What survived: the four research aims and their structure.
• What didn’t survive: the staging rationale—why that order, what the session actually argued, and where uncertainty remained.

The result was a clean artifact of what had been decided, but not a record of how the decision was reached.

The Two Layers of LLM Sessions

1. Artifacts – files created, decisions made, nodes extracted, code written. These persist because they are stored in a file system, vault, or repository.
2. Deliberative Reasoning – the why: framing choices, rejected alternatives, staging logic, and live uncertainties. This layer makes the artifact navigable and understandable for future review.

Current LLM tooling provides robust infrastructure for artifacts but nothing for preserving the deliberative reasoning layer. When a session ends (or compresses), the deliberative layer evaporates. Each compression pass discards branching paths and retains only conclusions, causing the staging rationale and alternative framings to disappear.

The Gap in Existing Tooling

• ECP nodes capture processed research.
• Session notes capture operational events.

Neither captures the reasoning that connects one piece of work to the next. The infrastructure for this “deliberative” layer simply does not exist yet.

The /mark Skill

• Purpose: Capture deliberative moments during working sessions, not after the fact.
• Design criteria:
  • Not a session note (which records what happened).
  • Not an extracted research node (which records processed content).
  • Focus on the layer between—design decisions in flight, research framings being constructed, and the logic linking current work to future work.

Anchor Format

A session anchor is a markdown file that:

1. Records what was decided.
2. Captures why (the argument as it stood at capture time).
3. Lists what remained uncertain.

These three elements are sufficient to make a decision navigable two sessions later.

Failure When Capturing From a Summary

The same session that produced the /mark skill also produced the first anchor. By the time I invoked /mark, a compaction event had already occurred, so the research framing existed only in the pre‑compression transcript, while I was working from the summary.

• Structure captured: four research aims with framing questions.
• Reasoning missed: staging rationale, ordering justification, and lingering uncertainties.

The structure survived double compression; the reasoning did not. This illustrates that compression systematically preserves conclusions while discarding the reasoning trace.

Lessons Learned

• Capture must happen live, before any compression, at the moment of insight.
• Running /mark retroactively yields the same “artifact‑without‑argument” problem that session notes produce.
• After a session ends, the deliberative reasoning is already gone; only the artifact layer remains.

Recommendations for LLM Practitioners

1. Identify the layer you need to capture.

   • Session notes → what happened.
   • Summaries → what was decided.
   • Deliberative anchors → why the decision was made.

2. Capture in real time. Use tools like /mark during the live session, before any compression occurs.

3. Validate your capture. After a session, try to reconstruct the reasoning from your notes. If you cannot, the reasoning has evaporated.

By integrating real‑time capture of deliberative reasoning, we can preserve not just the artifacts of LLM sessions but also the arguments that make those artifacts meaningful.