Reading
Jaquatech's open-source Claude Memento Prompt converts Microsoft Research's Memento technology into directly usable system prompts. Through chunked reasoning and memory compression mechanisms, Claude can handle ultra-long texts and complex multi-step tasks, extending effective reasoning length without model fine-tuning.
Section 01
This article introduces Jaquatech's open-source Claude Memento Prompt solution, which converts Microsoft Research's Memento technology into directly usable system prompts. Through chunked reasoning and memory compression mechanisms, Claude can handle ultra-long texts and complex multi-step tasks without fine-tuning, extending effective reasoning length.
Section 02
Although current mainstream large language models have large context windows, in practical applications, reasoning quality decreases significantly as the number of dialogue turns increases and text length grows. "Mid-context loss", cumulative errors in multi-step reasoning, and context truncation are major challenges for developers. Microsoft Research proposed the Memento technology to address this issue, and Jaquatech has transformed it into an open-source implementation solution.
Section 03
Memento is a technology developed by Microsoft Research to extend the effective reasoning length of large models. Core idea: Split long reasoning chains into discrete reasoning blocks, each generating a compact summary (memento fragment), discard detailed content, and only retain memento fragments for subsequent reasoning. It simulates human thinking by preserving key conclusions and intermediate results to advance thinking.
Section 04
The original Memento requires special tokens and KV-cache masking to be implemented in engines like vLLM. Jaquatech has transformed it into a pure prompt engineering solution through carefully designed system prompts, which any Claude user can use immediately. Core structural elements: reasoning blocks (sub-problem reasoning within tags), memento fragments (dense summaries within tags), progressive mechanism (expanding subsequent reasoning based on memento fragments), and final synthesis (generating answers by summarizing memento fragments).
Section 05
The project provides multiple usage methods:
/memento:memento <task>;system parameter to the Anthropic API.Section 06
Memento Prompt performs well in the following scenarios:
Section 07
Value: Zero-cost extension of reasoning capabilities—no need for model fine-tuning, special hardware, or modifying the inference engine; only through prompt engineering to enhance Claude's ability to handle complex tasks. Limitations:
Section 08
This project is implemented based on Microsoft Research's Memento paper, using the CC0 public domain license. The original authors have no affiliation with Microsoft; it is an independent community implementation. Claude Memento Prompt transforms cutting-edge research into an immediately usable tool, which is worth trying for developers dealing with complex multi-step reasoning tasks. As large model applications expand, context management technology will become more important.
Keep going with more reads from the same topic.
Splinter is a minimalist, high-performance key-value (KV) and vector storage system enabling zero-latency inter-process communication via shared memory and atomic operations. With only 766 lines of core code, it supports millions of operations per second and 768-dimensional vector storage, offering a new architectural approach for local LLM inference and data-intensive applications.
Folkering OS is the world's first AI-native bare-metal operating system, entirely written in Rust no_std without relying on Linux, POSIX, or libc. It can generate commands from scratch, compile them into WASM, and run them in 10 seconds, achieving true self-evolution.
This project explores how to use Large Language Models (LLMs) to detect logical vulnerabilities in Ethereum smart contracts, especially business logic flaws that are difficult for traditional static analysis tools to capture, such as economic manipulation and execution order errors.
Cross-generational hardware testing based on NVIDIA A10G and H100 GPUs, comparing and analyzing the throughput, latency, and concurrent scalability of three mainstream LLM inference frameworks. SGLang achieves a 3.4x performance advantage over vLLM on H100, while Ollama faces architectural bottlenecks in high-concurrency scenarios.