Documentation

Feedback loop

The Feedback Loop

The feedback loop is the axis that lets Mnemix learn from outcomes. A caller agent reports what actually happened after it acted; Mnemix audits the context read and adjusts what it surfaces next time. This is what keeps the brain self-correcting instead of frozen.

The feedback record (beta)

A feedback record links a served memory packet to an outcome verdict and, when available, a correction. The public voice API does not expose this beta intake surface yet; it remains an internal substrate primitive until it graduates into the frozen public API.

| Feedback verdict | Meaning | | :--- | :--- | | helpful | The surfaced context was right and useful. | | wrong | The context was incorrect. | | stale | The context was outdated. | | missing | The context the agent needed wasn't surfaced. |

What happens downstream

Today (as built): a feedback row does two things, immediately —

  1. Audits the read. The original context read's audit record (context_audit) is updated with the outcome, so every packet Mnemix served has a verdict attached to it.
  2. Adjusts memory confidence. The referenced memory objects get an immediate confidence adjustment — wrong/stale signals push the offending chunks down in future retrieval; helpful reinforces them.

On the roadmap (not yet live): nightly reranker calibration from accumulated verdicts, grader-threshold tuning on miscalibrations, and automatic locked-fact supersession proposals from wrong + correction pairs. Those are designed but not shipped — this page will say so when they are.

Closing the loop on coding (the dogfood case)

The clearest example is Mnemix governing its own construction. Internally, a GitHub Action can hand the merged PR outcome to the private feedback intake:

on:
  pull_request:
    types: [closed]
jobs:
  feedback:
    if: github.event.pull_request.merged == true
    runs-on: ubuntu-latest
    steps:
      - run: echo "Internal beta: send the merged PR outcome to the private feedback intake"

The public docs intentionally stop at the handoff point because this intake surface is spec-frozen but not publicly callable yet. When a merged PR later breaks staging or needs a revert, the internal workflow records that verdict and optional correction so the corpus learns which kinds of changes drift.

Why this is a primitive, not a feature

Without feedback, governance is static: the rules never improve, retrieval never adapts. The feedback loop is what turns a set of rules into a system that gets better the more it's used — while still never deleting history.

See also