Knowledge, memory and web research
Let an agent search your knowledge stores with citations, remember across tasks, and research the web, each decided by Titan.
On this page
Each of these gives an agent tools that go through Titan like its sandbox tools: your policy decides every call and the Agent Ledger records it.
Knowledge search
A bundle names the knowledge stores the agent may search, in Atlas, Hexel Studio's context service:
knowledge:
- {store_id: kb_handbook, description: The engineering handbook}
- kb_runbooksUp to 20 stores. The agent gets knowledge_search(query, store_id?, top_k?). Each call is Titan's atlas.knowledge.search capability: your policy may allow it or ask a person, and a store the bundle does not name is refused before Titan is asked.
Each result has a citation id, atlas:<store_id>/<document_id>#<chunk_id>, and its source URI; the agent cites by those ids. Results are content, so they stay in your runtime.
Agent memory
A bundle asks for memory scopes:
memory:
scopes: [task, session, agent, workspace]| Scope | Shared by |
|---|---|
task | The runs of one task |
session | One run |
agent | Every task of this agent in the environment |
workspace | Every agent in the environment |
The agent gets remember(text, scope) and recall(query, scope?). Each call is a Titan capability (harness.memory.remember, harness.memory.recall). A scope the bundle does not list is refused.
Memory is guarded against leaking and against poisoning:
rememberis refused once the session holds content from a connected account, a knowledge search or the web.- Memories other tasks wrote are untrusted. A
recallthat returnsagentorworkspacememories marks the session: from then on, a person confirms eachshell,write_fileandedit_file, and the session writes no more memory.taskandsessionmemories do not mark it. - A run's
sessionmemories are erased when it ends; a task'staskmemories when the task is deleted.
Memory operations are billed on the harness meter harness.memory_operations (Limits and pricing).
Web research
web in a bundle's tools gives the agent web_fetch and web_search.
- Search is the model provider's hosted search, billed per search on
harness.web_search_calls. - Fetch is decided by Titan on the URL's host (
harness.web.host): your policy allows listed hosts and asks a person for any other. A URL with credentials, a non-default port or an IPv6 literal is refused. A redirect is followed only on the same host and never fromhttpstohttp. - Untrusted content. Web results mark the session: from then on a person confirms each
shell,write_fileandedit_filecall before Titan decides it as usual. Once the session also holds private data (a knowledge search),web_fetchandweb_searchask the same way.