Sillage

Your language model forgets everything. Sillage gives it a memory — and a

sillage (n., French) — the trace left behind by something that has passed:

a ship's wake, a scent in a room. What a model keeps of what it read.

time — no gradients, no fine-tuning, no growing index . One Hebbian matrix

that consolidates by surprise, and a rank-16 adapter on the readout. Four

mechanisms, five papers, one command-line tool . Everything runs on a

draft costs half the perplexity it would have cost on Monday (10.68 -> 5.39),

and the model completes a sentence it can only know from what it read.

sillage read notes.md # memorize it (CPU: ~8 min per 10k tokens)

read preprint_v1.txt: 5859 tokens in 4.9 min | PPL 12.14 -> 11.83 (adapter) -> 11.71 (+memory)

read preprint_v2.md: 11041 tokens in 9.5 min | PPL 10.68 -> 9.82 (adapter) -> 5.39 (+memory)

memory consolidated and saved (16900 tokens lifetime, 9883 cold grams, 194 passages indexed).

The memory lives in ./.sillage and survives restarts — it survived a power

cut during development. Its size is fixed the day you start: 7.4 MB with

document or ten thousand. It never learns from its own generations ,

8 minutes per 10k tokens with the default Qwen3-0.6B, about 2 with

--model gpt2 — which is 4x faster but English-only. index and ask need

Verified across three architectures (GPT-2, Qwen3, GPT-NeoX). Four things to