Salience & Confidence

Every memory carries two meta-cognitive scores that influence how it is handled: salience (emotional/motivational significance) and confidence (epistemic certainty). Together, they determine which memories are protected from pruning and which should be flagged as uncertain during recall.

Salience (0.0 - 1.0)

Salience captures how emotionally or motivationally significant a memory is. High-salience memories represent things that deeply matter — critical decisions, hard-won lessons, or personally important events.

Salience Levels

RangeLevelDescription
0.0 - 0.3LowRoutine information, minor observations
0.3 - 0.5ModerateUseful knowledge, standard decisions
0.5 - 0.7HighImportant decisions, significant learnings
0.7 - 1.0CriticalCareer-defining decisions, hard-won lessons, core values

Protection From Pruning

Memories with salience >= 0.7 are never auto-pruned. Even if their strength decays below the archive threshold (0.1), they will not be moved to _archived/ during the sleep cycle. They must be explicitly removed via /brain:forget (or /brain:forget --deep for full forensic erasure).

This models how emotionally significant memories persist in the human brain long after neutral memories fade.

Consolidation Anchoring

During consolidation (a phase of the sleep cycle), the memory with the highest salience in a group serves as the anchor. Its framing, key details, and perspective take priority in the synthesized result. Lower-salience memories in the group contribute supplementary details.

Scoring Impact

Salience contributes 0.08 (8%) of the scoring formula:

score = 0.38 * relevance
      + 0.18 * decayed_strength
      + 0.08 * recency_bonus
      + 0.14 * spreading_bonus
      + 0.14 * context_match
      + 0.08 * salience          ← this factor

While 8% may seem modest, for memories with similar relevance and strength scores, salience can be the tiebreaker that determines which memory surfaces.

Synaptic Homeostasis Re-Boost

During the Synaptic Homeostasis phase of sleep, high-salience memories receive selective re-boosts after the global downscaling. This means important memories are preserved even when the overall memory system is being normalized.

Confidence (0.0 - 1.0)

Confidence tracks how epistemically certain a memory is — how much you should trust it. Not all memories are equally reliable. Some are well-established facts, while others are tentative observations or secondhand information.

Confidence Levels

RangeLevelDescription
0.0 - 0.3LowUnverified, tentative, or potentially outdated
0.3 - 0.5ModerateLikely correct but not confirmed
0.5 - 0.7StandardBased on direct experience or reliable sources
0.7 - 1.0HighVerified, well-established, or authoritative

How Confidence Is Set

Confidence is initially set at encoding time based on the quality of the source:

  • Direct experience — You did it yourself, saw the result → 0.8-1.0
  • Reliable documentation — Official docs, peer-reviewed sources → 0.7-0.9
  • Team knowledge — Information from colleagues → 0.5-0.7
  • Secondhand or inferred — Heard from someone, read in a forum → 0.3-0.5
  • Speculation or assumption — Educated guess, untested hypothesis → 0.1-0.3

How Confidence Changes

Confidence is not static. It changes through several mechanisms:

Knowledge Propagation (during sleep):

  • When new evidence contradicts a stored memory, its confidence is reduced by -0.20
  • When new evidence validates a stored memory, its confidence is boosted by +0.10

Manual adjustment:

  • You can ask the agent to update a memory's confidence when you discover new information

Low-Confidence Flagging

During recall, memories with confidence < 0.5 that have been accessed frequently (access_count >= 3) are flagged:

⚠️ 2 frequently-used memories have low confidence — consider verifying

This warns you that you are relying on uncertain information. The agent may also add a caveat when referencing a low-confidence memory during a session.

warning

Low confidence does not prevent a memory from being recalled — it just adds a warning. This is by design: uncertain memories can still be useful, but you should be aware of their reliability.

Salience vs. Confidence

These two dimensions are independent:

CombinationMeaningExample
High salience, high confidenceCritical and reliableA verified architecture decision that shaped the entire system
High salience, low confidenceImportant but uncertainA hypothesis about a production bug that has not been confirmed
Low salience, high confidenceReliable but routineA well-documented API endpoint format
Low salience, low confidenceUncertain and unimportantA casual observation about a library version
tip

If you notice the brain status showing frequently-used memories with low confidence, review and either verify or discard them — the /brain:sleep cycle also surfaces these for reinforcement.