Skip to content

Memory and Learning

NataPulse uses memory to preserve useful context across time without treating every past conclusion as permanently true.

A completed research run can create a safe lesson containing product-relevant elements such as:

  • subject and horizon;
  • thesis or stance;
  • public confidence;
  • key evidence;
  • invalidation signals;
  • what to monitor;
  • later outcome status when available.

Internal traces, costs, secrets, raw error messages, and private administrative scores are excluded.

When a new investigation concerns the same or a closely related subject, the Research Supervisor can retrieve relevant lessons. Retrieval is scoped and ranked by relevance and recency.

Semantic retrieval can help find related concepts, while structured storage remains authoritative.

New research does not simply overwrite history. A newer lesson can supersede an older one while preserving the relationship between them. This makes it possible to inspect how the interpretation changed.

A prior research view can later be marked:

  • confirmed;
  • partially confirmed;
  • invalidated;
  • inconclusive.

Outcome tracking improves evaluation and helps future research distinguish reliable patterns from recurring errors.

Explicit feedback on answers and reports contributes another quality signal. Feedback is useful when it identifies unsupported claims, weak citations, missing counterevidence, or poor relevance.

Memory is not a guarantee that the system “learned the truth.” A stored lesson can reflect incomplete evidence or a mistaken conclusion. Future agents should compare it with current evidence rather than copying it.

Memory remains subject to workspace isolation, retention, privacy, redaction, and product publication rules. Private workspace information is not automatically converted into public NataPulse data.