Research log · 2026-09-05
Smaller context. Measured progress.
We are testing whether a shared dictionary can reduce repeated conversation context. We publish the overhead and the cases where compression does not help.
Looking for your own usage? Open your dashboard. Project owners can see recorded successful compressions in Operations.
Aster-1 · experimental dictionary codec
Repeated text becomes a short reference with a complete decoder. The software decoder restored every tested input exactly. This does not establish that an AI model will answer identically.
This experiment is separate from the current hosted compressor and npm release. The current compressor uses heuristics; test its results against your task requirements.
First benchmark
Synthetic fixtures, measured locally using gpt-tokenizer 3.4.0. Each turn includes the growing input history, decoder, dictionary, escaping, and JSON. Repeated handoffs deliberately favor dictionary compression. These are not customer savings, provider bills, or a forecast.
| Fixture | Encoding | Original tokens | With codec | Reduction |
|---|---|---|---|---|
| repeated handoff · 24 turns | cl100k | 24,324 | 10,673 | 56.12% |
| unique short messages · 4 turns | cl100k | 35 | 35 | 0% |
| marker heavy literals · 8 turns | cl100k | 2,672 | 2,672 | 0% |
| repeated handoff · 24 turns | o200k | 23,748 | 10,647 | 55.17% |
| unique short messages · 4 turns | o200k | 39 | 39 | 0% |
| marker heavy literals · 8 turns | o200k | 2,672 | 2,672 | 0% |
Excludes provider role framing, tool schemas, outputs, retries, latency, and cache pricing. These encodings do not establish token counts for Claude, MiMo, MiniMax, or Gemini. Model answer equivalence has not been established.
When does the dictionary pay off?
In the repeated-handoff fixture, turn 3 is the first profitable turn. Early inputs stay unchanged. Each bar below counts the full input for that individual turn, including the decoder again.
Turn 1: 82 → 82 tokens · unchanged
Turn 3: 244 → 222 tokens
Turn 8: 649 → 342 tokens
Turn 24: 1,945 → 726 tokens
Live model smoke test · MiMo
One synthetic development handoff, tested twice with mimo-v2.5: original input 715 tokens, encoded input 394 tokens, with 44 output tokens in each response. The model returned the same correct project facts, including a “do not publish” restriction, but wrapped both answers in Markdown. Both therefore failed the requested strict JSON format check.
This small test does not establish general answer equivalence. Cached tokens were present in the repeat; token reduction is not a measured bill reduction. No hidden assistant instructions or tools were compressed.
Next generations
- Aster-1 — literal dictionary · local prototype
- Exact software round trips, full payload accounting, and a bypass when overhead is larger than the saving.
- Lyra-2 — model evaluation · proposed
- Compare original and encoded prompts on held-out tasks. Count correct answers, failures, input and output tokens, retries, and latency.
- Orion-3 — session integration · proposed
- Preserve roles and tool boundaries, validate dictionary versions, and compare compression with provider caching. No launch date or savings target is promised.
Names identify research milestones, not new foundation models. A generation advances only with reviewable evidence.
Help choose the next test
Tell us which repeated workflows cost you tokens. Share a synthetic example without private data or credentials.
Send feedback