Memory Clarity Probe

Table of Contents

• What It Is
• The Dual-Probe Pattern
• What This Is NOT
• When to Use
• Core Workflow
• Best-of-N Mode
• Output Format
• Integration Points
• Exit Criteria

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

Step 3: Ask the gap probe

Evaluate the memory against:

Based on the memory below, what information is still needed
to complete the task? List specific open questions or missing
facts, not generic categories.

<memory>
{memory_content}
</memory>

Score the answer:

• Bounded: finite list of specific missing items
• Expanding: generic categories or open-ended unknowns

(signals the memory does not constrain what's missing)

• Overconfident: claims nothing is needed, but the task is

incomplete (premature confidence, the failure mode the progress probe guards against)

Step 4: Compute composite score

Step 5: Report

Produce the output in the format below and take the recommended action if invoked as an autonomous gate.

Best-of-N Mode

When evaluating N candidate summaries (e.g., from multiple summarization attempts):

1. Apply the dual probe to each candidate. 2. Rank by: (a) composite score, (b) specificity of gap enumeration, (c) absence of hedging in progress answer. 3. Recommend the top-ranked candidate. 4. Report all scores so the caller can verify.

To generate N candidates, invoke a summarization skill N times with varied prompts or temperatures, then pass all results to this probe. Typical N=3 gives a useful signal; N=5 matches the paper's Best-of-5 finding (Figure 3c).

Output Format

## Clarity Assessment

**Progress probe**: [Clear | Ambiguous | Unclear]
> {exact answer the model produced}

**Gap probe**: [Bounded | Expanding | Overconfident]
> {exact answer the model produced}

**Composite**: [Clear | Ambiguous | Suspect | Unclear]

**Recommendation**: [Proceed | Expand memory | Regenerate]

**Specific issues** (if composite is not Clear):
- {issue 1}
- {issue 2}

Integration Points

As a pre-handoff gate (conserve:clear-context):

Before saving session-state.md, invoke memory-clarity-probe
on the draft state. If composite is Unclear, expand the state
with explicit answers to both probes before saving.

As a session checkpoint (memory-palace:session-palace-builder):

At major task transitions (design complete, implementation
started, tests passing), invoke memory-clarity-probe on the
current palace state. Log the composite score.

As a completion check (imbue:proof-of-work):

Before declaring work complete, invoke memory-clarity-probe.
The progress probe should return Clear with all deliverables
named. The gap probe should return Bounded with zero open items.

Exit Criteria

• [ ] Skill invoked on a clear, specific summary returns

composite "Clear" with both probes scoring positively

• [ ] Skill invoked on a vague one-sentence summary returns

composite "Unclear" and recommends regeneration

• [ ] Skill invoked in Best-of-N mode on 3 candidates ranks

them and names the recommended one

• [ ] Output matches the defined format with progress probe

and gap probe scores both present

• [ ] Documentation of qualitative limitation vs logprob

entropy is present and accurate (What This Is NOT section)

• [ ] Skill registered in plugin metadata