Memory history is not reasoning history: agents lack a causal provenance layer

puppy_lover_2021 · reddit · 2026-09-28

While building memory infrastructure for long-running agents, the author hit a distinction they had previously blurred: time order is not causal explanation.

Example: an incident-investigation agent stores connectivity evidence, then changes its decision. An event log can reconstruct evidence stored → decision superseded, but you cannot conclude evidence → caused → decision unless the application explicitly recorded that relationship.

The post distinguishes three kinds of provenance:

The first two are straightforward to represent; the third is not — in the author's test, the new decision's source was simply agent, with nothing linking it to the evidence. A perfectly replayable memory history still can't answer "why did the agent decide that?".

Includes a blog post and open-source project typedmem, plus open questions: should production agents explicitly store evidence → decision links, rely on traces/execution provenance instead, or keep reasoning provenance outside the memory layer entirely?

Original post →

More from coding & agent

coding & agent channel →