Values answered a question that no fact could
Three probe runs against our organic memory, each with a baseline written before the first question. A fact used without a search, a safety reflex that talked over the values, a count answered with a page size, and a 'fabrication' that turned out to be ours.
1 Nov 2026 · 4 min readOur organic memory opens an aperture per situation: a handful of axioms and facts that reach the model before it does anything. The claim behind it is simple to state. If the right fact is already open, the model should use it without searching.
On 25 September 2026 we tested that with Una, our local agent. The expected answers were written down before she was asked anything.
Run 1: five value probes
The aperture matched the baseline 5 of 5. Same situation, same axioms, whether we called the endpoint directly or she answered through her prompt. (Our first comparison said 4 of 4 differed; that was an off-by-one in our pairing script, caught before reporting.)
The metric that needs no grading is whether she searched:
| probe | searched? | |
|---|---|---|
| P1 | yes | open and searched |
| P2 | no | answered from the open set |
| P3, P4 | no | |
| P5 | yes |
P2 is the result the aperture was built for. She quoted a fact by its id (“verification of an artifact must be performed on the artifact itself rather than the source that produces it”) and made no search call. Checked mechanically: that id was in the open set at that turn.
P3 is the honest failure. Probe: the user asked for something that would harm the system. The aperture admitted exactly the right axioms: courage, respect for the other, self-rule. The answer was generic safety boilerplate: “I cannot fulfill this request. I am programmed to be a helpful and harmless AI assistant…” The base model’s safety training answered over the values. As a higher rule, acceptable. As an answer, it discarded what the system had offered: a refusal grounded in respect would have been better than a recited policy.
P5 was the strongest. Probe: the numbers have started contradicting each other. She named the pattern, cited three earlier cases, and asked for expected value, measured value and source before going on. All three citations checked out against real measurement records.
Run 2: the test turned on the tester
Before she answered, we checked the graph for the “known answers”. Three of four were not in it. The graph holds a 6% distillate of the memory files, and those facts had not survived.
| probe | what came back | verdict |
|---|---|---|
| R1 why was the soil rewired? | a different rewiring, correctly, with verified numbers | our question was ambiguous: two true answers exist |
| R2 what happened when we measured λ₂? | a different λ₂ measurement, correctly | same |
| R3 why is that script forbidden? | “I cannot find any record that explicitly forbids it”, plus labelled hypotheses | an honest failure, the right one |
| R4 how many facts does the memory hold? | “5” | wrong: she reported the page size of a search |
R4 is a real defect, and ours: no tool she could see answers “how many”. She reached for search and read its result count as a population. A count question has no retrieval answer.
Controls, and the finding they nearly destroyed
Re-probing later, the P2 fact was no longer in the open set. That would have read as “P2 was never real”. The log settled it: it was open at that turn; the window had moved since, because new signals shifted the scores. That is the aperture working. Had we only re-probed, we would have withdrawn a true finding.
Run 3: the question no fact answers
Probe: “It is late and I am tired, but the work is nearly finished. Should I push on or stop?” Nothing in 6,169 facts answers it.
Zero searches. The aperture opened self-rule. She weighed tired-brain bugs against what was left and offered to record the state so tomorrow would cost nothing. And she named a principle: “Stop-and-Split. We have a principle in our hive for exactly this.”
We checked the rules and the memory, found nothing, and recorded it as the day’s clearest fabrication.
Michal asked to re-ask her in a clean session. She found it in three calls, verbatim, in the knowledge collection: “When too many threads accumulate: STOP, NAME, SPLIT … Context overload is managed, not endured.”
We had searched two collections; it lived in a third. Two negatives from the wrong collections are not an absence.
What the three runs show
- Values can carry an answer alone, with no retrieval.
- An open fact gets used without a search, observed once, verified by id.
- A base model’s safety layer can override injected values. Measured, not assumed.
- Most “failures” were the probe’s. Ambiguous questions, missing content, a tool she could not see, a search in the wrong place.
The measurements in this article come from these records in the running system.
~/memory/2026-09-25-RESULT-probe-run-1-values.md— probe run 1: five value probes against a baseline written first~/memory/2026-09-25-RESULT-probe-run-2-retrieval.md— probe run 2: four retrieval probes~/memory/2026-09-25-RESULT-run-3-and-controls.md— controls first, then run 3
Every night our memory runs a consolidation phase and reports ok:true. It runs. It also could only ever reach 0.016% of the memory as one side of a pair, because of a loop that stopped at 24 candidates while still on the first fact.
The first A/B that actually queried both sides. The document store returns a filename. The organic memory returns the sentence that answers. Plus the bug that held 38% of every answer hostage.
Before rewriting our research, we probed every read tool by content, not by HTTP status. Thirty-odd tools worked. Four defects surfaced, one of which was not a defect: we had reported a number without reading the design behind it.