Distributed Engrammics, preprint and reproducible code. Can a skill be transferred from one AI instance to another without fine-tuning, without backpropagation, and without updating model weights? This work treats a model's fast-weight state as a persistent, transferable object, an "engram", that can be written, read, moved between instances, forgotten, and governed. Fast weight programmers treat the fast-weight matrix as an ephemeral memory, reset every sequence and private to one network. We study the opposite: this matrix F as a persistent, transferable object exchanged between agents, with four operations (inscribe, read, transfer, forget) and a subspace governance rule, tested through six pre-specified falsifiable hypotheses. In a controlled associative-memory backend (60 seeds, paired-bootstrap 95% CIs, Holm–Bonferroni), a low-rank state delta extracted from one agent transfers a skill to another without any gradient step, separating from both a no-transfer and a norm-matched random control by ≈88 points, with a rank dose-response, preserved host skill, projection-based forgetting, and subspace-gated consent. We then run the identical protocol on real DeltaNet language-model backends (1.3B and 2.7B, via flash-linear-attention). The primary claim holds: a gradient-free engram added to a recipient's recurrent state transfers the skill (+54 to +66 points over both controls across the two sizes, p < 0.001), with a monotone rank dose-response, and subspace governance succeeds once the authorized subspace is read from the model's actual keys (+62 points at 1.3B). Two idealized properties degrade: superposition interferes with the host skill, and subtraction-forgetting damages it. We trace this to shared key directions and confirm it by making the skills use disjoint symbols. Beyond memorized lookups, the fast state can carry a generalizing rule: a vowel/consonant classifier engram, injected into another instance, classifies held-out letters above no-transfer, random, and label-swapped controls on DeltaNet, and largely replicates on a second architecture (RWKV-7, where it clearly beats the no-transfer and label-specificity controls). On RWKV-7-2.9B, which can induce a novel rule in context, the engram of a freshly learned Caesar shift transfers that non-trivial induced rule: a neutral recipient applies it to held-out letters at 0.44 versus 0.005 (random) and 0.08 (wrong-shift), p < 0.0001; this extends to a small family of induced rules (further shifts and a reflection, each beating a wrong-rule control), though specificity does not hold in a new digit domain. Most experiments use two instances of one frozen checkpoint, where moving a state is close to state interpolation and the genuinely new content is the algebra of governance and forgetting; our most novel result lifts that restriction. Between two distinct RWKV-7 checkpoints, naive injection fails at chance while a learned linear key alignment recovers it to the recipient's own ceiling (0.04 → 0.97): a closed-form, skill-independent map fit once per pair with both models frozen, so the per-skill transfer stays gradient-free. We position the method against activation-space task and function vectors and state-soup interpolation, and release a single reproducible harness whose backends are judged by identical criteria. What this record contains: the paper (LaTeX source + compiled PDF), a single reproducible harness with two backends (a controlled associative-memory model and real linear-attention language models), per-seed CSVs, and complete raw run logs. Code and full reproduction instructions: https://github.com/JacquesGariepy/engrammics Scope: the mechanism applies to linear-attention models with a single additive recurrent state (DeltaNet, RWKV-7), not standard softmax (GPT-like) transformers; a follow-up paper may explore extending it to those architectures.
Jacques Gariepy (Fri,) studied this question.