Configuration

All memory and knowledge settings are managed through the CrabShack dashboard. Changes take effect automatically — you don't need to restart anything.

To access settings, navigate to Memory or Knowledge Base in the sidebar and click the Settings button (gear icon).

Memory Settings

Setting	What It Controls	Default	Range
Auto-extract from conversations	Whether memories are automatically learned from agent sessions	On	On/Off
Inject memories into sessions	Whether relevant memories are included in new conversations	On	On/Off
Conversation summarization	Whether to generate rolling summaries of long conversations	Off	On/Off
Hot memories	How many memories are always included in every conversation, regardless of topic	20	1–100
Max memory context	Maximum amount of memory text injected per conversation (in characters)	8,000	1,000–50,000
Memory cap	Total number of memories stored before older ones are automatically archived	10,000	100–100,000

Recommendations

• Auto-extract and Inject memories should generally stay on — they're the core value of the memory system. Turn them off only if you want fully manual memory management.
• Hot memories — keep this between 10 and 30 for most use cases. Too many hot memories means the agent spends most of its context budget on memories rather than the conversation.
• Max memory context — the default of 8,000 characters works well for most conversations. Increase it if your agents handle complex, context-heavy tasks. Decrease it if conversations are short and focused.
• Memory cap — the default of 10,000 is generous for most installations. Only increase this if you run many agents with frequent conversations.
• Conversation summarization — enable this if your agents have long, detailed conversations and you want to preserve a condensed version. When enabled, you can select which model to use for summarization.

Knowledge Base Settings

Setting	What It Controls	Default	Range
Auto-inject knowledge into chat	Whether relevant document excerpts are automatically included in conversations	On	On/Off
OCR for images	Whether to extract text from images and scanned PDFs	Off	On/Off
Search precision	How closely a document must match to be included — lower is broader, higher is stricter	0.3	0.1–0.9
Chunks per query	How many document excerpts are returned per search	5	1–20
Max knowledge context	Maximum amount of document text injected per conversation (in characters)	12,000	1,000–50,000
Adjacent sections	How many surrounding sections to include with each match (for additional context)	1	0–2
Document limit	Maximum number of documents in the Knowledge Base	500	1–10,000
Storage limit	Maximum disk space for document storage (0 = automatic, uses 50% of free space)	Automatic	0–100,000 MB
Cloud sync interval	How often to check connected cloud drives for changes	Every 6 hours	1–168 hours

Recommendations

• Auto-inject should stay on for most use cases. If you have a very large document library and want agents to only search when explicitly asked, you can turn it off.
• OCR — enable only if you work with scanned documents or images containing text. It adds processing time for each document.
• Search precision — the default of 0.3 is a good starting point. If you're getting too many irrelevant results, raise it toward 0.5–0.6. If the agent is missing relevant documents, lower it toward 0.2.
• Chunks per query — increase this (to 8–10) if your documents are densely packed with relevant information. Keep it lower (3–5) if you want concise, focused excerpts.
• Adjacent sections — 1 is a good default. Set to 0 if your documents have very distinct, self-contained sections. Set to 2 if sections are short and context from neighboring sections is frequently needed.
• Document limit — increase beyond 500 only if you have a large library to index. More documents means slightly longer search times.
• Storage limit — the automatic setting works well for most deployments. Set a manual limit if you need to reserve disk space for other purposes.
• Cloud sync interval — every 6 hours works for most teams. Set to hourly if your documents change frequently and freshness is critical.

Common Configurations

Customer Support Agent

A support agent that needs quick access to policies and procedures:

Setting	Recommended Value	Why
Hot memories	15–20	Keep key customer context always available
Auto-inject knowledge	On	Agents should reference support docs automatically
Search precision	0.25–0.35	Cast a wider net to find relevant policy sections
Chunks per query	5–8	Include enough context to answer detailed questions

Research Assistant

An agent that works with many documents and performs deep analysis:

Setting	Recommended Value	Why
Max knowledge context	20,000–30,000	Allow more document content per conversation
Chunks per query	10–15	Return more excerpts for thorough analysis
Search precision	0.2–0.3	Broader search catches more potentially relevant content
Adjacent sections	2	Include more surrounding context for each match

Small Business (Few Documents)

A small team with a handful of key documents:

Setting	Recommended Value	Why
All defaults	—	The defaults work well for small document libraries
Document limit	500 (default)	More than enough for most small businesses
Cloud sync interval	24 hours	If documents rarely change, less frequent syncing is fine

Enterprise (Large Document Library)

An organization with hundreds or thousands of documents:

Setting	Recommended Value	Why
Document limit	2,000–10,000	Accommodate a large library
Search precision	0.35–0.5	Stricter matching avoids noise from a large corpus
OCR	On (if applicable)	Ensure scanned documents are searchable
Cloud sync interval	1–6 hours	Keep up with active document management
Storage limit	Set manually	Control disk usage in shared environments

tip

When in doubt, start with defaults and adjust based on what you observe. Use the Retrieval Test tool to see the effect of settings changes before they impact real conversations — it's the fastest way to tune your configuration.