Memory

The memory is post-cortex, the one module that runs in every organization and every daemon and cannot be turned off. Every other module keeps what it knows in it.

Why it works the way it does: Graph, Search, Conversations, Cards. The pages in the app: Memory.

What it keeps

• Workspaces and sessions: a workspace is one project; a session is one line of work in it, with a role (primary, related, dependency). See Workspaces.
• Records: what an agent learned, each of a type (a question and its answer, a decision, a problem solved, a code change, a requirement, a definition), with the entities it names and how they relate. A record that a newer one replaces stays in the history, marked as ended.
• The graph: the entities the records name and the typed relations between them, so a question about one thing also finds what depends on it.
• Conversations: the turns of an agent’s session, captured as they happen, searchable like records.
• Cards: the user card (one per person) and the project card (one per workspace), short notes put in front of every session.
• Documents: what other modules keep in a workspace (a board’s tasks, a workflow, a skill, a mail template), versioned; only the module that declares a kind writes it.

Secrets and personal data are masked before anything is stored: Security and data.

Search

Every search chooses a mode:

Mode	What it does
fast	words and meaning together, no reranking
local	reranked by a model on your own daemon
accurate	reranked by Voyage AI in a workspace that turned it on; elsewhere, on a daemon, by its local model

The cloud runs no local reranker: there local, and accurate in a workspace without Voyage, answer as fast does, and every answer says whether it was reranked. The numbers behind each mode: Search.

Settings

• In the cloud: each workspace’s name, description, and whether accurate searches may use Voyage: Settings.
• On a daemon: search, redaction, conversation capture and storage in daemon.toml: Configure.

Tools

Tool	What
session	make, list, load and search sessions; find or make a project’s workspace and session
update_conversation_context	write records, one or many, with their entities and relations
semantic_search	search a session, a workspace or everything, in one of the three modes
assemble_context	everything for one task in one call: the cards, the entities named, the best records, within a token budget
query_conversation_context	a session’s decisions, open questions, an entity’s neighbourhood, one record whole
get_structured_summary	a session’s summary: decisions, insights, entities
manage_workspace	workspaces, their sessions, and the cards
manage_entity	delete an entity or a single record
manage_messages	a session’s conversation turns: append and page
manage_document	the documents modules keep, with their history
admin	health, embeddings of a session, a checkpoint

Every argument of every tool: Tools reference. The same over HTTP: REST API.