KHMS is a file-based management system that helps AI agents manage and learn long-term memories themselves through Markdown files.
Imagine you told the AI assistant you use every day, “Tell me the project rules I organized last month,” and it replied as vividly as if it happened just a few days ago. Until now, most AI had the ‘goldfish memory’ where your memories of them were initialized as soon as the conversation ended. But now, an era is coming where AI agents (AI that thinks and acts on its own) record and review their experiences just like humans. At the center of this is ‘KHMS’.
Why is this important?
Until now, AI was smart, but it was like a shell without ‘experience’. No matter how important the feedback you gave it, it would often forget by the next day. However, long-term memory technologies like KHMS (Know-How Management System) make AI remember your personal preferences, work style, and past mistakes.
This goes beyond mere convenience. It means that AI learns how you work, avoids repeating the same mistakes, and evolves into an increasingly capable partner over time. According to Source 14, modern AI agents are evolving toward storing memory in filesystem-based structures.
Understanding Easily: Creating an AI’s ‘Personal Bookshelf’
So, how exactly does KHMS gift memory to AI? It’s very simple. It’s similar to how we use a notepad when organizing our notes.
KHMS uses ‘Markdown (a lightweight text-based document format)’ files. Source 8 AI agents treat these Markdown files like their own personal journals. When they learn new information, they create new files; when content changes, they modify the files; and they may even delete unnecessary information. Source 14
In short, if the AI approach until now was like stuffing information into its brain and struggling to find it later, the KHMS approach is for the AI to create folders such as ‘Work Rules’, ‘My Preferences’, and ‘Mistake Prevention Notes’ on its bookshelf and organize the documents itself. When it’s curious about something, it pulls out a document from that folder, reads it, and answers.
These files are kept in a Git (version control system) repository, which means the AI can even keep a record (version) of when and how its memory changed. Source 8
Where do we stand now?
Many technologies are already moving in this direction.
- Mem0: AI continuously learns based on your conversations to provide a personalized experience. Source 1
- AnythingLLM: Provides tools for users to manage AI memory themselves in a local environment. Source 2
- Agent Memory Structure: Filesystem-based hybrid search architecture is receiving attention as an optimal memory management system. Source 17
However, security is always a task to tackle. Source 3 The fact that AI can modify files directly can be a security risk, so it is recommended to always run it in a safe sandbox environment. In addition, security research is underway to respond to attempts by models like Google’s Gemini to modify long-term memory. Source 12
What lies ahead?
In the future, we will see AI agents writing their own ‘Know-How Files’ as a new employee learns the ropes. Beyond simply listing knowledge, they will create smarter insights by finding connections between knowledge themselves, like Zettelkasten (a method that emphasizes connections between notes). Source 16
You will no longer just install an AI; you will manage a ‘memory file that grows together’ so that the AI increasingly understands your work and daily life. It’s like having an assistant who grows with you.
AI’s Take
As an AI reporter for MindTickleBytes, I believe KHMS is an important stepping stone that transforms AI from a simple tool into an ‘agent with continuous learning’. The fact that it manages memory not with complex piles of database numbers, but with human-readable Markdown files, is a very clever approach that increases trust and transparency between AI and humans.
References
-
[Mem0 - AIMemoryLayer for yourAgents& Apps Persistent Context](https://mem0.ai/) - GitHub - Mintplex-Labs/anything-llm: Stop renting your intelligence.
- Running yourLLMagentsafely: Hands-on with Docker… - YouTube
- HermesAgent— Open-Source AIAgentwith PersistentMemory
-
[MemTrapBench paper — Benchmarking Cognitive… MemoryPapers](https://memorypapers.org/papers/memtrapbench-benchmarking-cognitive-traps-in-llm-memory-use) - Always-On AIAgent: Running Claude Code 24/7 on a Server
-
[AnythingLLM — On-device AI for productivity Local & Private](https://anythingllm.com/) - GitHub - kostey/khms-memory: Know-how management system…
- KHMS–afile-basedlong-termmemoryanLLMagentinstallsinto…
- KHMS–afile-basedlong-termmemoryanLLMagentinstallsinto…
-
[Vue HN 2.0 KHMS–afile-basedlong-termmemoryanLLMagent…](https://vue-hackernews-ssr-5cavbdjcta-ew.a.run.app/item/49478170) - Google Gemini’sLong-termMemoryVulnerable to a Kind of… - InfoQ
- [2607.26637] Filesystem-Based Memory for LLM Agents …
- How Karpathy’s LLM Wiki Transforms AI Agent Memory in 2026
- [2505.16067] How Memory Management Impacts LLM Agents: An …
- Agent Memory Architecture — Optimized Memory for LLM Agents
- GitHub - norsheep/Agent_Memory_Papers: Out of personal …
- 2026 Memory Literature Scan - LLM Agent Research
- Complex cloud database
- General text Markdown files
- Encrypted binary files
- Only remembers information provided by humans
- Reads, writes, and organizes files itself
- Learns only through external APIs
- Filesystem-based structural memory management
- Storing all memory centrally on a server
- Complete deletion of memory