A proactive system needs memory, but not every stored value has the same epistemic status. nano-os-agent journals attempted experiments. Application adapters maintain observations and confirmed facts. PicoClaw explains released evidence. Chat history is not a truth source for board state.
Different records answer different questions
| Class | Example | Permitted statement |
|---|---|---|
| Observation | A calibrated sensor read 31.2 C at a timestamp |
The sensor reported this value under the recorded conditions |
| Released experiment result | A focus-control task passed all declared checks | This run passed its checks |
| Human-confirmed fact | The operator confirmed a treatment or sample identity | The event was reported and confirmed |
| Public source | A paper or manual with URL and retrieval time | The source states or supports a candidate method |
| Proposal | Inspect a sample or test a sensor placement | This is a suggested next action, not an executed fact |
Decide what may be used as current evidence
Field-related nano-os-agent results use a narrow release rule:
- Passing and internally consistent: retain as
observed_successwith the scope of that run. - Failed, partial, blocked, or step-inconsistent: retain the result as diagnostic evidence but exclude its measurements from operational advice.
- Missing current application state: create an internal refresh or investigation request rather than inventing a current value.
This rule limits what may influence downstream explanations. It does not erase failure; failure remains valuable diagnostic evidence.
Keep the path back to the source
Every durable item should answer:
- What subject or sample did this concern?
- When was it observed or reported?
- Which skill, task, sensor, person, or source produced it?
- Which units, model version, and configuration applied?
- Did release checks pass?
- Was the item observed, confirmed, inferred, or proposed?
Without these fields, a numeric value may be precise but scientifically weak.
Time order does not prove cause
Suppose a farmer confirms an operation and a later observation is clean. The agent may store:
{
"relation": "observed_association",
"operation_id": 41,
"outcome_id": 57,
"causal_claim": false
}
The sequence can guide a future question, but it does not establish that the operation caused the outcome. Stronger claims require controls, replication, or another defensible design.
Old measurements need an age label
Evidence also has a useful lifetime. A fresh dashboard, sensor snapshot, or task result can answer a current question. Yesterday's value may remain valid history, but it should not be silently reused as today's state. Application adapters therefore define freshness and regeneration rules.