Write-Side Enrichment: Automatic Tags and Category

What it fills in

Write-side enrichment fills in the metadata a caller didn’t supply on neuron memory add — tags and category. It fills only unset fields; anything passed explicitly (--tags, --category) is honored untouched and never overridden. It hangs off NeuronMemory.transact, the single seam every write in neuron routes through, so enrichment applies uniformly whether the write came from the CLI or from neuron scan.

Centroid matching, not free-form generation

Tags are selected by centroid cosine similarity against a closed vocabulary — every tag declared in neuron.yaml, plus every store tag already carried by at least three entries — with no model call involved. A tag’s centroid is the normalized mean embedding of the entries that already carry it; selecting a tag is really “which existing cluster does this new entry’s content resemble.” Category is inferred the same way, by default, against the store’s own category centroids. A live A/B test measured this centroid approach against a model-based alternative that could read each category’s description field as an instruction: centroid won 9 of 9 corpus questions, the model 1 of 9 — most of the model’s answers weren’t even a declared category. Centroid is the default for both, because of that result and because it removes the last model call from the write path.

The model cannot mint a tag

Both tags and category are drawn from a closed, declared set — the model (when a categoryStrategy: model opt-in is configured) cannot invent a new tag or category from embedding proximity alone. Minting a new tag stays a human act, which keeps a store’s tag vocabulary from fragmenting into one-off near-synonyms.

Importance is never inferred

Importance was a third enrichment target through two earlier release candidates. A benchmark measured the shipped 0.5B model’s discrimination between deliberately unambiguous critical and trivial entries at -0.5 and +0.167 across two runs — noise, not signal — including at least one case where the model rated a note about irreversible production data loss as importance 1. That inference path was removed outright rather than kept as a disabled default. An omitted --importance takes the column default of 3.

Source: CONTEXT.md “write-side enrichment”, ADR 0010 — Guardrails for the 0.5B Model’s New Jobs.