Feedback that trains recall

PLUR’s recall ranking improves as you use it. Every plur_feedback call nudges an engram’s activation; over time, the engrams that help you stay sharp, and the ones that don’t fade. This page is the practical guide to making that loop work.

The two signals

plur feedback ENG-2026-0525-001 positive
plur feedback ENG-2026-0525-001 negative

Positive says: this engram was injected, and it actually helped. The next task you face, similar engrams should surface earlier.

Negative says: this engram was injected, and it didn’t help (or was off-topic). Next time, surface it later — or not at all.

When to give positive feedback

• The agent visibly used the engram (cited it, followed its rule).
• You’d be annoyed if it didn’t surface for the same kind of task again.
• It saved you from re-explaining a convention.

Don’t over-reward. If an engram surfaces ten times and helps three times, that’s worth three positive signals — not ten.

When to give negative feedback

• The engram surfaced for a task it has nothing to do with.
• It contradicts a more recent decision you’ve made.
• The agent followed the engram and produced a worse result.

Strong negative signals (3+ in a short window) move an engram toward dormant, where it’s no longer eligible for default injection.

Implicit feedback (no work from you)

The adapter’s hook layer reads implicit signals you give without realising:

• You corrected the agent after a turn → the engrams injected for that turn get a negative implicit signal.
• You moved on without comment → mild positive signal.
• You explicitly thanked the agent → positive signal.

Implicit signals are weaker than explicit ones — your conscious plur feedback call carries more weight than a heuristic guess. But the implicit channel matters at volume.

Spotting engrams that need feedback

After a session, scan what got injected:

plur timeline --channel session --limit 1 --json | jq '.engrams_injected'

For each:

• Was it useful? → positive
• Was it noise? → negative
• Was it neither — just irrelevant? → don’t bother

The decay-vs-strengthen balance

Engrams accrue feedback the same way they accrue access. Activation is built up from:

• Time since last access (decays).
• Positive feedback (strengthens).
• Negative feedback (decays faster).
• Recall hits (strengthens slightly).
• Injection use (strengthens slightly).
• Pinning (bypasses decay entirely).

The math is in Activation & decay. The rule of thumb: feedback faster than time decays.

Anti-patterns

• Mass positive at the end of every session — degrades signal, every engram looks important.
• Negative for every engram you didn’t read — many engrams help silently; you don’t always see their effect.
• Re-saving instead of feedback-ing — if an engram is mostly right, give feedback. Don’t create ENG-…-002 saying nearly the same thing.

Feedback in Enterprise

Per-user feedback is recorded server-side. Aggregate signals roll up to the engram itself — if an engram has 47 hits and 13 misses, all injections weigh that ratio. Personal feedback weights more heavily for your recalls, but team-wide signal informs ranking for everyone.

The Insights dashboard at /admin/insights surfaces engrams with high injection counts and low feedback ratios — strong candidates for retirement.

Where the loop actually closes

Three sessions in, give it a glance:

plur list --status active --sort activation

The top engrams should be the rules and conventions that actually matter in your day-to-day. If they’re not, your feedback isn’t being given (or is being given to the wrong engrams). Adjust and try again.

The loop works. It just needs a few hundred signals to learn what you care about.