Memento

Paper (PDF) | OpenMementos Dataset

Memento extends the effective output length of large language models by splitting chain-of-thought reasoning into blocks and summaries (memento). After each reasoning block, the model generates a short summary, then the block content is evicted from the KV cache. The model continues from the summary with a shorter context, enabling more reasoning within a fixed context window.

Special tokens

Token	Purpose
<think> / </think>	Reasoning wrapper
<|block_start|> / <|block_end|>	Reasoning block boundaries
<|summary_start|> / <|summary_end|>	Summary (memento block) boundaries

Block structure: <|block_start|> reasoning <|block_end|> <|summary_start|> summary <|summary_end|>

Repository layout

Directory	Description
data/	Data pipeline — converts raw CoT traces into the Memento format (block boundaries + summaries) for SFT training
vllm/	vLLM overlay — adds KV cache block masking to stock vLLM for efficient Memento inference

Quick start

Data pipeline

Convert chain-of-thought traces into Memento training data:

pip install -r data/requirements.txt
export OPENAI_API_KEY=sk-...   # or any OpenAI-compatible provider

cd data/pipeline
python run_full_pipeline.py \
    --input ../examples/example_trace.jsonl \
    --output-dir output/ \
    --model gpt-4o \
    --limit 1

See data/README.md for full documentation.

vLLM inference with block masking

• Step 1: Set Up the Environment Build a customized vllm with block masking support:

  pip install vllm==0.13.0
  cd vllm
  bash install_overlay.sh

• Step 2: Serve a Memento Model with KV Cache Compaction To expose the model through an API-compatible server, run:

  python -m vllm.entrypoints.openai.api_server \
      --model /path/to/memento-checkpoint \
      --served-model-name memento \
      --port 8010 \
      --max-model-len 32768 \
      --gpu-memory-utilization 0.9 \
      --trust-remote-code \
      --chat-template chat_templates/memento_nosys.jinja \
      --block-masking-config '{
          "enable": true,
          "keep_last_n_blocks": 0,
          "mask_delimiters": false,
          "compact_on_summary_end": true,
          "require_assistant_section": true,
          "debug": true
      }'

  See vllm/README.md for full documentation, including API usage and alternative setup options.

License

This project is licensed under the MIT License.