Memory should return evidence, not confidence
Noru Pro memory is retrieval over local captures, not training. It can suggest a small number of prior observations, preserve their provenance, and return nothing when nothing earns its place.
On this page
“Memory” is an attractive word because it promises continuity. It is also an easy way to blur several very different mechanisms.
Noru does not train a model on your screen. It does not convert repeated exposure into private knowledge inside an agent. It stores local observations and retrieves them when the current task makes earlier evidence relevant.
Retrieval is a selection problem
Exact text search is useful when the agent knows the string it needs. Contextual recall is different: it tries to surface earlier observations related to the fresh capture and the agent's request, even when the wording is not identical.
The safe output is not an unbounded replay of history. It is a small set of candidates with capture identity, time, matched text, and a reason the item was suggested. The agent can then fetch a specific capture and inspect the evidence directly.
A hint is not a fact
Similarity says that two observations may belong together. It does not say which one is current, which one is correct, or whether the resemblance matters to the task. That is why historical hints remain visibly historical and why fresh evidence outranks remembered context for questions about the present.
The distinction also protects the image itself. A retrieval hint can stay compact rather than silently injecting several old screenshots into the agent's context. Pixels arrive only when the agent explicitly retrieves a chosen capture.
Abstention is part of the feature
A memory system that always returns something will eventually return a persuasive distraction. Sometimes current evidence is sufficient. Sometimes the history contains no answer. Sometimes every candidate is too weak to justify the context it would consume.
In those cases, no relevant prior evidence is a useful result. It keeps uncertainty visible and gives the agent less irrelevant material to rationalize.
Provenance makes memory auditable
Every returned candidate points back to a stored capture rather than a detached summary. The timestamp, source, and capture identity let the agent inspect where a remembered fragment came from. If two observations conflict, the conflict can remain visible instead of being averaged into confidence.
The Noru decision
The purpose of local recall is not to make the agent feel certain. It is to make relevant prior evidence available with less searching and fewer hidden assumptions.
Useful memory does not say “trust me.” It says “this earlier observation may matter; here is where it came from.”