Persistent, on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
Persistent, on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
Persistent, on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
Persistent, on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
An on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
Reason-why
After experimenting with several system prompts for a while, I eventually moved toward collecting the various guidelines in a RAG system. Both solutions have specific advantages and disadvantages:
System prompts are passed at the start of a conversation. As the conversation progresses, they are sometimes forgotten — especially when the instructions describe very specific use cases.
Embeddings are delivered context-sensitively in the respective turn. Although the content is tailor-made — provided all parameters of the RAG pipeline are well calibrated — the mode of delivery has two critical disadvantages: 1. The push principle does not create the best reception conditions, only the second-best. (Even if "spoon-feeding" seems convenient, it is not always accepted by the recipient.) 2. The communication is one-sided: the agent itself has no opportunity to document and record what it has learned.
The best experience I've had so far has been with my tool: Playbook. It combines all the advantages and avoids all the disadvantages mentioned:
Playbook
An on-demand knowledge base for LM Studio agents. Write instructions, skills, and notes in Markdown or plain text, let the agent find, read, write, and edit them — without polluting the context window on every turn.
Reason-why
After experimenting with several system prompts for a while, I eventually moved toward collecting the various guidelines in a RAG system. Both solutions have specific advantages and disadvantages:
System prompts are passed at the start of a conversation. As the conversation progresses, they are sometimes forgotten — especially when the instructions describe very specific use cases.
Embeddings are delivered context-sensitively in the respective turn. Although the content is tailor-made — provided all parameters of the RAG pipeline are well calibrated — the mode of delivery has two critical disadvantages: 1. The push principle does not create the best reception conditions, only the second-best. (Even if "spoon-feeding" seems convenient, it is not always accepted by the recipient.) 2. The communication is one-sided: the agent itself has no opportunity to document and record what it has learned.
The best experience I've had so far has been with my tool: Playbook. It combines all the advantages and avoids all the disadvantages mentioned:
The pull principle ensures optimal reception conditions: the agent receives the information it needs exactly when it actively asks for it.
Beyond the ability to maintain the knowledge base editorially and curate its contents professionally, agents can add their own notes and update outdated information.
The sometimes very specific information contained in the system does not bleed into "other" topics and content of conversations. The tool can remain "always on" without increasing the token load (aside from the tool descriptions).
Even better might be fine-tuning the agent models — which is, however, comparatively complex and harder to keep up to date.
How It Works
Playbook is like the opposite of a classical RAG pipeline: nothing gets injected automatically. Instead, the agent uses tools when it decides the knowledge base is relevant.
Two-phase retrieval keeps context usage lean:
recall — keyword + semantic search, returns metadata only (title, filename, tags, score). The context cost is negligible.
read — fetches the full text of one specific document, on demand.
Write access is equally explicit:
Tool
What it does
memorize
Write or overwrite a document (title + tags + Markdown body)
rewrite
Surgical text replacement — oldText must match exactly once
forget
Delete a document permanently
Setup
Install the plugin.
Load an embedding model in LM Studio. The default is text-embedding-nomic-embed-text-v1.5 — swapping in any English or multilingual model works.
In the plugin settings, set Playbook Directory to the absolute path of a folder where your Markdown and plain text files should live. This is the only required setting.
That's it.
If no embedding model is configured (or LM Studio can't reach one), search falls back to BM25 keyword matching automatically.
File Format
Documents are Markdown (.md) or plain text (.txt) files with optional YAML frontmatter:
---title: "My Note"tags: ["example", "setup"]created: "2026-04-07T10:00:00.000Z"updated: "2026-04-07T10:00:00.000Z"---Your content here.
Frontmatter is written and maintained automatically by the memorize tool. Plain text files without frontmatter are indexed too — the filename is used as the title. You can also create and edit files manually in any editor — the plugin detects changes via file watcher and re-indexes within seconds.
Configuration
Setting
Default
Description
Playbook Directory
~/Documents/Playbook
Absolute path to the folder with your MD/TXT files
Embedding Model
text-embedding-nomic-embed-text-v1.5
Leave empty for BM25-only search
LM Studio Base URL
http://127.0.0.1:1234
LM Studio local API URL
Max Recall Documents
5
Maximum results per recall call
Min Recall Score
60
Minimum relevance score (0–100) to appear in results
Technical Requirements
LM Studio with plugin runner support
An embedding model loaded in LM Studio (optional, for semantic search)
Companion plugins
draw-things-chat — text2image / image2image / video generation, with Vision Promotion so the agent can see and evaluate its own output
analyse-image — reads generation metadata from generated images for easy re-generation. Provides masks for agentic inpainting or zoom-in tasks with process-image based on object detection.
draw-things-index — metadata query across previously generated images
user-docs — Personal Knowledge Management supported by vision-capable agents. Fully integrates all playbook capabilities and lets you re-use all your instructions, skills, and notes created with playbook
Disclaimer
‼️ This plugin is a research prototype. ‼️
License
MIT
The pull principle ensures optimal reception conditions: the agent receives the information it needs exactly when it actively asks for it.
Beyond the ability to maintain the knowledge base editorially and curate its contents professionally, agents can add their own notes and update outdated information.
The sometimes very specific information contained in the system does not bleed into "other" topics and content of conversations. The tool can remain "always on" without increasing the token load (aside from the tool descriptions).
Even better might be fine-tuning the agent models — which is, however, comparatively complex and harder to keep up to date.
How It Works
Playbook is like the opposite of a classical RAG pipeline: nothing gets injected automatically. Instead, the agent uses tools when it decides the knowledge base is relevant.
Two-phase retrieval keeps context usage lean:
recall — keyword + semantic search, returns metadata only (title, filename, tags, score). The context cost is negligible.
read — fetches the full text of one specific document, on demand.
Write access is equally explicit:
Tool
What it does
memorize
Write or overwrite a document (title + tags + Markdown body)
rewrite
Surgical text replacement — oldText must match exactly once
forget
Delete a document permanently
Setup
Install the plugin.
Load an embedding model in LM Studio. The default is text-embedding-nomic-embed-text-v1.5 — swapping in any English or multilingual model works.
In the plugin settings, set Playbook Directory to the absolute path of a folder where your Markdown and plain text files should live. This is the only required setting.
That's it.
If no embedding model is configured (or LM Studio can't reach one), search falls back to BM25 keyword matching automatically.
File Format
Documents are Markdown (.md) or plain text (.txt) files with optional YAML frontmatter:
---title: "My Note"tags: ["example", "setup"]created: "2026-04-07T10:00:00.000Z"updated: "2026-04-07T10:00:00.000Z"---Your content here.
Frontmatter is written and maintained automatically by the memorize tool. Plain text files without frontmatter are indexed too — the filename is used as the title. You can also create and edit files manually in any editor — the plugin detects changes via file watcher and re-indexes within seconds.
Configuration
Setting
Default
Description
Playbook Directory
~/Documents/Playbook
Absolute path to the folder with your MD/TXT files
Embedding Model
text-embedding-nomic-embed-text-v1.5
Leave empty for BM25-only search
LM Studio Base URL
http://127.0.0.1:1234
LM Studio local API URL
Max Recall Documents
5
Maximum results per recall call
Min Recall Score
60
Minimum relevance score (0–100) to appear in results
Technical Requirements
LM Studio with plugin runner support
An embedding model loaded in LM Studio (optional, for semantic search)
Companion plugins
draw-things-chat — text2image / image2image / video generation, with Vision Promotion so the agent can see and evaluate its own output
analyse-image — reads generation metadata from generated images for easy re-generation. Provides masks for agentic inpainting or zoom-in tasks with process-image based on object detection.
draw-things-index — metadata query across previously generated images
user-docs — Personal Knowledge Management supported by vision-capable agents. Fully integrates all playbook capabilities and lets you re-use all your instructions, skills, and notes created with playbook