Ordo Research Lab — rewritable memory for frozen LLMs
Ordo is an independent research lab led by Russel Gavery (Gavrilov Ruslan, @8hrsk). Its goal is a locally
hosted assistant that accepts a new version of a documentation set as a diff instead of being retrained, and
that is measured honestly against the retrieval pipelines it would have to replace.
In plain terms
A language model's knowledge is frozen at training time. When a library ships a new version or an internal
policy changes, the model is confidently wrong, and the sanctioned fix is to train again — which is expensive
and destructive: an equal-capacity fine-tuning baseline corrupts 8.4% of unrelated facts on a single edit.
Ordo keeps the facts in a rewritable store the frozen model reads from, so an update is a small write rather
than a training run. Two consequences follow. A company can keep a local assistant current on its own private
documentation without a vendor round-trip. And a model or agent vendor could ship knowledge updates on their
own cadence, separately from the model — a coding assistant carrying current knowledge of popular libraries
in the model itself, instead of every customer running retrieval or MCP servers on their side.
Ordo-M — rewritable external memory
A value table addressed from text is attached to a frozen model, leaving the base weights bit-identical. A
point edit causes 0.0% collateral damage, against 8.4% for an equal-capacity LoRA. On a real code domain the
memory passed the bar it set for itself against retrieval, 68.1% against a 67.8% threshold, but only after
the corpus was rewritten into the form the questions ask about: fed raw documentation text it scores 54.1%,
and a deliberately wrong address scores the same as the right one. On a second, harder domain the bar of 71.2%
was cleared at 71.69% by two value tables at different depths sharing a single address — and that verdict has
since been read more strictly. Measured in the chat framing a product actually uses, the memory reads 64.10%,
but so does everything it is compared against: hybrid retrieval falls from 92.31% to 78.30% and the oracle arm
from 93.29% to 78.50% even with the right record quoted verbatim, so the bar itself moves to 64.50%. Both
margins, +0.49 and −0.39, are smaller than the disagreement between the two modules that compute the two
halves of the criterion — 6 and 2 questions of 507. The quantity reported in its place is the one that
resolves: paired on the same questions, the memory is 20.32 and 20.91 points behind hybrid retrieval in the
bare framing and 14.00 and 14.40 behind in the product's, p < 0.0001. What the two sites buy was renamed
too: at equal flow they add no knowledge (+0.00 and −0.35 points) and cost less damage (a single site costs
+1.68 and +1.38 more). The mechanism does transfer to prose, +3.17 points at p = 0.023, on one seed. The
damage side is solved separately: a gate with no trainable parameters, open where an answer is written and
closed on ordinary prose, removes 96% of the text damage for a cost of 2.0 points of knowledge that is not
statistically significant. Incremental updates hold over twenty real releases at 4.5 times less compute than
a full retrain. The track is written up as a preprint at
alphaXiv.
Methods, measurements and negative results are published in English and Russian; source code and evaluation
harnesses are closed. See @ordoproject and
github.com/Ordo-Project. Contact:
founder@ordo-project.com.