Enrichment

Enrichment is the background process that turns raw events into the summaries and search index that make memory useful.

What it does

A worker watches for new events and schedules enrichment jobs:

• Transcript chunks that group raw events near their source text.
• Turn summaries from individual exchanges.
• Trace summaries for tool-use sequences.
• Session and chat summaries rolled up from their children.
• Hourly and daily summaries that roll activity up by time window.

A language model produces the summaries, routed through the same proxy, so the model and credentials are resolved per account. The default enrichment model is a fast, inexpensive Gemini model. You can override it per account under Settings -> Processing.

Indexing

After summaries are written, the indexing pass embeds them and pushes them to the search index (see Hybrid search).

Cost controls

Enrichment is designed to be economical:

• Events are de-duplicated, so replayed history is not re-summarized, and a summary job that already succeeded is not run again.
• The worker applies quota-aware backoff: when a provider rate-limits or hits a daily cap, it pauses rather than burning spend on retries.
• You can disable enrichment per account if you only want capture and raw recall. Jobs for an account with enrichment off stay queued and resume if you turn it back on.

Configuring

Choose your enrichment and embedding models, and toggle enrichment or embedding, from Settings -> Processing in the dashboard.