01
Use is the signal
Not age, not size, not a schedule. A lesson earns a shorter form only when it was loaded, actually used, and the work succeeded — which is proof it contained slack. Served but ignored earns nothing.
For Claude Code and Codex. It learns what you had to correct, so you never have to say it twice — and the context it costs you goes down the longer you use it.
Nobody wrote that lesson. You corrected the agent once, in the course of normal work, and ROSE kept the conclusion instead of the transcript — 4,200 tokens of fumbling reduced to 260.
Then it earned its way shorter twice: used, worked, shortened. A second lesson arrived from another repo and the two became one. Store went up, tokens loaded went 260 → 118.
Every other memory moves those together — remember more, load more. Pulling them apart is the entire product.
Append-only stores keep everything and grow until retrieval is guesswork and the context bill dwarfs the task it was meant to help with.
Summarisers stay cheap by throwing things away, and what they throw away first is the exact port, the one flag that mattered. The loss is invisible until something breaks.
Not age, not size, not a schedule. A lesson earns a shorter form only when it was loaded, actually used, and the work succeeded — which is proof it contained slack. Served but ignored earns nothing.
Compression adds a level above. The verbose original sits underneath it forever, so "shorter" is never a bet you can lose. The worst case is that you walk back down.
A shorter lesson must declare exactly what it dropped, as discrete claims. That turns walking back down from a search into a lookup — and it is why the short form is safe to load first.
This is the question that sinks most compression schemes. Compression normally destroys the very information you would need to invert it, so recovering the right detail becomes a blind walk down a tree.
ROSE makes recording the loss mandatory. Every compressed lesson carries a manifest of the claims it removed and which node still holds each one — so "which detail do I need back?" is a short ranked list, not an exploration.
| When | What runs | You wait |
|---|---|---|
| You submit a prompt | Matching lessons are injected | ~0.1s, and no model call at all until the store outgrows the budget |
| A substantial turn ends | A reflector reads the session off-thread | nothing — it is a detached process |
| The session ends | Lessons minted, compression attempted, replayed against past work | nothing |
| Once a day, with new evidence | Lessons used together earn a shared parent | nothing |
Everything expensive is detached, so the only cost on your critical path is the injected text itself — and that is the number the whole design drives downwards.
An injection is never silent. You see
⋯ ROSE · 2 lessons · 512 tok — Retry policy, Cache TTLs
above your prompt, naming what was loaded. A memory system that edits
your prompts without saying so is one you cannot notice going wrong.
Install once, then work normally. ROSE runs in hooks: it injects on prompt submit, reflects off-thread when a turn ends, mints and compresses when a session ends, and consolidates once a day. Every expensive step is detached, so nothing blocks your turn.
Python 3.10+, at least one of claude or codex,
and zero third-party Python dependencies. The store is plain files in
your repo — nothing is sent anywhere.
uv tool install rose-memory rose install --scope user
/plugin marketplace add yiheinchai/rose /plugin install rose@rose
rose status # what it knows, what it cost rose tree # the graph, by level rose recall --prompt "..." # what it would inject
ROSE is young and its claims are falsifiable, which is the point.
rose eval replays held-out work against every level of a
lesson, with a control arm and blind grading, so a compression that
stopped working shows up as a number rather than a feeling.
Precision is the weakest number in the system today. It injects
lessons that turn out not to matter, and that is measured rather
than hidden — rose status prints it.
There is no second look mid-turn when the agent discovers it needs more, and it cannot search the store itself.
The reflector still misses lessons a person would catch. When it does, that is a defect to file, not a fact of life.