An organic memory returns the statement, not the file
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.
1 Nov 2026 · 2 min readOur hive has two long-term memories. The crystal is a knowledge graph over documents. Phyto is an organic memory: about 7,000 atomic facts with growth, decay and hormone-like signals, opened per situation through what we call the aperture. The botany is deliberate: trees keep a memory of past stress without a brain 1, and a living plant’s physiology can itself serve as a computing reservoir 2.
The first comparison script said phyto lost. It had never sent phyto the query.
The real A/B: 10 questions, both sides asked
| measure | value |
|---|---|
| facts returned per query | 13.0 on average |
| of which axioms | 5.0 |
| same for every query | 6 facts (46%) |
| specific to the query | 7 facts (54%) |
| latency p50 | 159 ms (146–249) |
4 of the 5 axioms rotate with the question. Only one, sapere aude, is unconditional.
The difference is the unit of retrieval
| question | crystal returns | phyto returns |
|---|---|---|
| stress as friction | a rule’s title | “Stress signals should function as cognitive friction, making the process harder and requiring more energy” |
| why did package.json get corrupted | an audit report’s name | “the shell tool’s working-directory fallback, which defaults to the home folder” |
| which model does Una run on | a project brief’s name | “Una is a local Gemma-4 model, chat on :8080, embeddings on :8081, vision via mmproj” |
The crystal returns a pointer: fetch the document, then read it. Phyto returns the statement: one call, about 30 tokens.
The defect: five sticky facts
Five facts appeared in all ten answers, whatever was asked. They held 5 of 8 non-axiom slots, 38% of every aperture.
The mechanism was one line: the reinforcement score was added to situational similarity instead of gated by it. Similarity sat between 0.73 and 0.75, and a reinforcement score of 1.48 won every time. The organism was remembering what it had thought about most, not what it was asked.
What we claim, and what we do not
CLAIM situational retrieval returns the statement, not the document:
one call, ~30 tokens, p50 159 ms
CLAIM 54% of the context is query-specific
NOT CLAIMED "beats RAG": no ranking benchmark against labelled ground truth
existed when this was measured
This measures the shape of the answer, not its correctness. A ranking benchmark with labelled queries now exists for the main search tool. Running phyto through the same benchmark is the next step.
- P. Krokene, M. H. Mageroy. Tree Memories: How Can Trees Remember Without a Brain?. Frontiers for Young Minds 12:1400253 (2024)saved pagecorpus: ~/phyto
- O. Pieters, T. De Swaef, M. Stock, F. wyffels. Leveraging plant physiological dynamics using physical reservoir computing. Scientific Reports 12, 12594 (2022)originalcorpus: ~/phyto
The measurements in this article come from these records in the running system.
~/memory/2026-09-29-MEASURED-phyto-vs-crystal-first-real-ab.md— MEASURED: phyto vs crystal, the first A/B that actually queries both sides
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.
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.
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.