knowledge

Knowledge

The Knowledge section lets you build structured, persistent information stores that agents can draw upon during conversations. It is separate from conversation history — knowledge is available across all conversations for the agents you configure it for.

Accessing Knowledge

Go to Settings → Knowledge in the top navigation.

Knowledge Types

Ariadne provides four types of knowledge entries:

1. Dynamic Knowledge

Dynamic Knowledge entries are facts or pieces of information that an agent actively keeps up to date. You define what the fact should contain; the agent periodically refreshes it using its available tools.

Examples:

• "Current EUR/USD exchange rate" — the agent queries a financial tool and stores the latest rate
• "Status of our CI/CD pipeline" — updated from a monitoring tool
• "Today's top headlines" — refreshed from a web search

Dynamic knowledge is ideal for information that changes over time and should always be current when an agent references it.

Creating Dynamic Knowledge:

1. Click Add Dynamic Knowledge
2. Give it a name and a description of what it should contain
3. The agent will populate and maintain it automatically

2. Agent Skills

Agent Skills are Markdown documents that give an agent specific behavioural instructions or background knowledge. Think of them as an extended system prompt stored externally.

Examples:

• A communication style guide ("Always respond formally; use British English")
• A process definition ("When asked to review a PR, follow these steps: ...")
• Domain knowledge ("Our product pricing structure is as follows: ...")

Skills are attached to an agent in Settings → Agents and are injected into the agent's context when relevant.

Creating an Agent Skill:

1. Click Add Agent Skill
2. Enter a title and write the skill content in Markdown
3. Assign it to one or more agents from the agent settings page

Skills are retrieved automatically when they're relevant to a conversation (by semantic similarity), so you don't have to attach every skill to every message — the agent pulls in the ones that fit what you're working on.

Self-Improving Skills

Agents also learn and refine their own skills over time:

• As you work — after a conversation, an agent may capture a reusable procedure it just followed as a new skill, or add a bullet to an existing one. This is additive, so nothing you wrote is overwritten.
• Overnight — a nightly optimisation pass tidies the skill set: refining wording, merging overlap, and pruning skills that are never used (skills that have been used are kept).

Skill Version History

Every change to a skill — whether you made it or the agent did — is snapshotted automatically. On each skill card:

1. Click History to expand its version list, showing each snapshot's change type, source (manual, immediate-extraction, or nightly-optimisation), reason, and timestamp.
2. Click Restore on any snapshot to roll the skill back to that version.

This means automatic edits are never destructive — you can always see what changed and undo it.

3. Documents

Documents are files associated with a conversation or project. They are stored as blobs and can be:

• Searched by the agent using semantic search
• Returned chunk-by-chunk for large files

Documents are typically attached via conversation file uploads (see Chat & Conversations), but can also be managed directly from the Knowledge page.

4. Key–Value Pairs (KVP)

KVP entries are simple named values available to agents at runtime. They are useful for:

• Configuration values agents need to reference ("JIRA project code: PROJ")
• Preferences ("preferred coding language: Python")
• Identifiers ("customer Salesforce ID: 00123")

Creating a KVP:

1. Click Add KVP
2. Enter a key name and a value
3. Select which agent(s) can access it

Assigning Knowledge to Agents

Knowledge entries are assigned per agent. Go to Settings → Agents, select an agent, and configure which skills and KVPs it has access to. Dynamic knowledge entries are associated with the agent that manages them.

Use Cases

Tips

• Keep Agent Skills focused — one document per topic is easier to maintain than one large catch-all document
• Use Dynamic Knowledge sparingly for values that truly need to be current; static facts are better as KVPs or Skills
• Documents benefit from good titles and summaries so semantic search can find them accurately