Kimi K2.5 System Prompt + Tools Schema + Context/Memory Artifacts

A research-oriented repository of extracted artifacts from Moonshot AI's Kimi K2.5 inference-time behavior: the observed system prompt, tool schemas, and context / memory assembly format.

---

Overview

Kimi K2.5 is Moonshot AI's latest large language model featuring a Mixture-of-Experts (MoE) architecture with 1T total parameters and a 256K context window.

Key Topics Covered

• Kimi K2.5 system prompt and operational boundaries
• Kimi K2.5 tool use / function calling schemas (8 tool categories)
• Kimi K2.5 memory system (memory_space_edits) and context management
• Citation formatting, file system constraints, and step limits

---

What This Repo Contains

This repo packages three primary artifacts plus analysis:

.
├── kimi-k2.5-system-prompt-and-tools/
│   ├── kimi_k2.5_system_prompt.md   # Extracted system prompt text
│   ├── tools.json                   # Tool schema / function-calling spec
│   └── context.txt                  # Context + memory block assembly format
├── ANALYSIS.md                      # Architecture breakdown / commentary
├── README.md
└── LICENSE

Use cases: alignment research, prompt engineering education, agent framework replication, and studies of tool-use + memory behaviors.

---

Quickstart

1) Browse the artifacts locally

git clone https://github.com/dnnyngyen/kimi-k2.5-prompts-tools.git
cd kimi-k2.5-prompts-tools

Then open:

• kimi-k2.5-system-prompt-and-tools/kimi_k2.5_system_prompt.md
• kimi-k2.5-system-prompt-and-tools/tools.json
• kimi-k2.5-system-prompt-and-tools/context.txt
• ANALYSIS.md

---

Tooling Overview (High Level)

Kimi K2.5 exposes a structured tool-calling interface that enables web access, data retrieval, image search, Python execution, and persistent memory modification. The core tool families are:

• Web Tools — general-purpose internet retrieval

  • web_search: query search engines
  • web_open_url: fetch and read webpage content

• Image Tools — multimodal search

  • search_image_by_text: find images using textual prompts
  • search_image_by_image: find visually similar images

• Datasource Tools — structured dataset access

  • get_data_source_desc: inspect available datasources
  • get_data_source: query datasource contents

• Python Sandbox

  • ipython: execute Python for analysis, computation, and chart generation (with output restrictions enforced by Kimi)

• Long-Term Memory Tools

  • memory_space_edits: add, replace, or delete memory entries across named memory spaces

See kimi-k2.5-system-prompt-and-tools/tools.json for the complete schema and argument structure.

---

Provenance / Methodology

Artifacts were documented through inference-time introspection and guided prompting, using Kimi K2.5’s own self-reporting behavior. No unauthorized access and no binary reverse engineering were used.

• Original extraction conversation: https://www.kimi.com/share/19c003f5-acb2-838b-8000-00006aa45d9b

Note: These behaviors and schemas may change between releases. Treat this repo as a snapshot.

---

Disclaimer

This project is not affiliated with or endorsed by Moonshot AI. This repository is provided for educational and research purposes only. Please comply with applicable laws and terms of service when using these artifacts.

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE.

---

Keywords: Kimi K2.5, system prompt, tool schema, function calling, agentic architecture, memory mechanism, context assembly, prompt engineering, AI alignment, mechanistic interpretability.