Memory Clarity Probe

Memory Clarity Probe

Assess whether a memory, summary, or session state retains enough task information to guide future reasoning.

What It Is

A quality gate for any memory or summary, based on the dual-probe pattern from MMPO (arXiv:2605.30159, Liu et al. 2026). The probe asks two anchor questions against the current memory and evaluates whether the answers are confident and complete:

1. Progress probe: "Based on current memory, what is the current task progress?"
2. Gap probe: "Based on current memory, what information is still needed?"

A clear memory answers the progress probe with specific, verifiable state (not vague placeholders) and enumerates bounded, concrete unknowns on the gap probe. An ambiguous memory produces hedging on the progress probe and open-ended uncertainty on the gap probe.

The Dual-Probe Pattern

The two probes target different failure modes:

• Confident-wrong: the model has a wrong but confident belief about task state. The gap probe alone misses this. The model claims it has enough. The progress probe catches it: if the stated progress contradicts known facts, the memory has drifted.
• Uncertain-incomplete: the model is uncertain about where the task stands. Both probes surface this: the progress answer hedges and the gap answer lists open-ended unknowns.

The MMPO paper's ablation (Table 4) shows progress+gap outperforms gap-only across all context lengths. Use both probes.

What This Is NOT

This skill implements a qualitative clarity assessment. It does not compute the token-level predictive entropy (Belief Entropy, Eq. 5 in MMPO) that the paper uses for RL training. Night-market has no access to the model's internal log-probabilities.

The paper's Table 6 shows that qualitative probing (labeled "direct-answer entropy", r=0.54) is weaker than true entropy (r=0.68), and can encourage premature confidence. Use this probe as a necessary quality check, not a sufficient one.

When To Use

• Before conserve:clear-context hands off to a continuation agent
• At session checkpoints in memory-palace:session-palace-builder
• Before committing a summary to a knowledge palace via memory-palace:knowledge-intake
• Before imbue:proof-of-work declares work complete
• When evaluating multiple candidate summaries (Best-of-N mode)

When NOT to Use

• As a substitute for actually reading the task requirements
• To validate factual correctness (the probe tests clarity, not truth)
• When the memory is trivially short (under 100 tokens: read it)

Core Workflow

Step 1: Receive the memory

Accept the memory or summary as input. Sources:

• The current session-state.md (from clear-context)
• A palace room's content (from session-palace-builder)
• A knowledge digest (from knowledge-intake)
• Inline text provided by the caller

Step 2: Ask the progress probe

Evaluate the memory against:

Based on the memory below, what is the current task progress?
Describe specifically what has been completed and what state
the task is in right now.

<memory>
{memory_content}
</memory>

Score the answer:

• Clear: specific completed steps, concrete current state, no hedging ("I think", "probably", "it seems")
• Ambiguous: some specifics but with hedging or gaps
• Unclear: vague ("some work was done"), generic, or empty