Best Practices for Administrators

Structuring Company Knowledge

1. Static Knowledge: Upload documents such as HR policies, IT manuals, or general FAQs into namespaces under vector.global.knowledge.*. Agents of all employees can access this knowledge for reading.
2. Business Collaboration: Use vector.global.business.* for company-wide knowledge collections, projects, and billing, where authorized users are permitted to make additions.
3. Thematic Separation: Use the hierarchy consistently (e.g., vector.global.privat.recipes vs. vector.global.business.marketing). This allows agents to be given targeted access only to the “domains” relevant to their role.

Configuring topK Correctly

The topK value determines how many memory hits are loaded into a run’s context at most.

• Too high → The context fills with irrelevant entries; the LLM loses focus
• Too low → Important information is missing from the context
• Recommendation: Start with topK = 5; increase to 10 for agents with complex domain knowledge

Granting Namespaces to the Tool Search

What the LLM can search with memory-search is listed solely in the “Tool-Only Read Namespaces” field — the “Read” field feeds automatic injection alone. An ingested knowledge base listed only under “Read” is out of the tool’s reach: the agent searches, finds nothing, and will likely answer from its world knowledge instead of your manual.

Rule of thumb: everything the agent should consult on request — manuals, documentation, project knowledge — belongs under “Tool-Only Read Namespaces”. Everything it should know unprompted — preferences, conversation notes — belongs under “Read”. Both is possible, but then requires an entry in both fields.

When the LLM searches without naming namespaces, everything under “Tool-Only Read Namespaces” is searched. That is the normal case — it does not have to guess the right namespace.

Phrasing the Query in Whole Sentences

Memory search is semantic: the meaning of the query is compared against the meaning of the stored passages. A single keyword carries too little meaning to land above the minimum relevance reliably.

Measured against an ingested manual: the query Timer reached a relevance of 0.35 and therefore returned nothing at the default threshold of 0.4. The same question asked as a whole sentence reached 0.51 and hit the right manual page.

The tool description now explicitly instructs the LLM to pass the user’s question in full and in their own words rather than shortening it to a keyword. The same hint is worth adding to your own task contexts when an agent conspicuously often finds nothing.

Edge case — exact identifiers: For strings such as E0.0 or TV00, semantic search is the wrong instrument — such codes carry almost no meaning that could be translated into a vector. The only remedy is to surround the question with context (“How is the TON timer programmed in Web@SPS?”).

Using Auto-Write Deliberately

allowWrite determines whether the system automatically writes to the memory namespace after each run.

• allowWrite = true only for agents that actually generate new knowledge (e.g., analysis agents, research agents)
• allowWrite = false for pure execution agents (sending emails, reading calendars) — these produce no insights worth storing permanently
• allowToolWrite = true only when the agent should actively and deliberately build up knowledge

Sub-agents: By default, each sub-agent writes automatically after its run if its policy has allowWrite = true. To prevent this (e.g., because the sub-agent output is only an intermediate result), set allowWrite: false in the sub-agent memory policy.

Monitoring & Security

• Dashboard: Regularly check the dashboard under “Memory & Audit.” A high number of warnings in the last 24h indicates misconfigured agents or attempted unauthorized access.
• Audit Log: Use the namespace filter in the audit log to specifically search for access to sensitive areas (e.g., global.business.billing).
• Ranking Rules: Use the “Namespace Rules Editor” to control the importance of certain sources. For example, vector.global.knowledge.* namespaces can receive a bonus so that verified company knowledge is preferred over fleeting chat notes.