Core Concepts

Kiomon is built around a few simple ideas. Once you understand them, everything else — the extension, the dashboard, the MCP tools, the API — falls into place.

The library is a corpus of documents

A document is any piece of saved content: a web page, an article, a note, a highlight, or a synced file from Notion, Google Drive, or Confluence.

• Each document has a title, an excerpt, tags, a source, and a body.
• Kiomon handles large content transparently — saves are instant and retrieval never slows down, no matter the size.
• Every document is auto-tagged and analyzed for the entities, claims, and domain it covers — this is what makes search and retrieval work well.

Documents live in a workspace. Documents in different workspaces are completely isolated.

Workspaces are isolation boundaries

A workspace is a named, isolated collection of documents and memories (e.g. “Personal”, “Work”, “Agent Handbook”). Everything you save, search, or ask about happens inside the context of a workspace.

• The browser extension lets you pick the workspace for each save (or create a new one on the fly).
• Agents discover workspaces via list_workspaces and declare one in every write.
• Free accounts get 2 workspaces; Pro gets 20.

Memories are agent-written knowledge

Where documents are imported content, memories are agent-authored knowledge: the facts, rules, preferences, and events your agents learn while working with you. Instead of agents dumping notes into a generic store, memories are first-class citizens with types and a lifecycle.

There are four memory kinds:

Kind	Friendly name	What it captures	Example
reference	Files	Imported/saved content and documents	A saved Stripe API page
semantic	Facts	Knowledge, facts, relationships	“The user’s product is a SaaS called Kiomon”
procedural	Routines	How-to rules, triggers, preferences	“Always answer with numbered citations”
episodic	Moments	Events that happened, with a timestamp	“Deployed v1.2 to prod on March 4”

Memories carry metadata that makes them useful to agents: entity, claim, and domain fields turn a blob of text into a searchable, graphable, consolidatable knowledge item.

The review-inbox flow

Anything an agent writes lands in your dashboard Review Inbox as a pending_approval draft. Drafts are not searched, surfaced, or embedded into retrieval until you approve them. Approve → the memory becomes active and joins your search and graph. Reject → it’s archived.

This is the guardrail that keeps your agent’s memory yours: nothing the agent writes becomes retrievable without your approval.

Lifecycle controls

• Pin — immune to decay, always ranks top.
• Archive — soft hide from retrieval, easily restored.
• Forget — permanent purge.
• Rate — +1 useful / −1 not useful; feedback feeds strength.
• Decay — unpinned memories slowly lose strength over time unless they’re retrieved or rated.
• Contradiction — when two memories contradict, the newer one reduces the older one’s confidence.

See Memories & Review Inbox for the full picture.

Retrieval is hybrid

Kiomon gives you two retrieval surfaces:

1. Search — hybrid: exact keyword matching fused with semantic (meaning-based) matching, so you find things by phrase and by concept.
2. Chat / RAG — grounded Q&A over your saved documents and memories, with numbered inline citations and a strict “answer only from the library” rule.

When retrieval surfaces a memory, Kiomon notes the usage — the more a memory is used, the stronger it gets.

The knowledge graph connects everything

Kiomon’s auto-linker (“the Librarian”) scans the library as documents arrive. For each new document it finds the content most similar to it and proposes link suggestions with similarity scores.

Approve a suggestion in the dashboard and it becomes a permanent, bidirectional link in your knowledge graph — visualizing how your whole library clusters. An interactive graph view renders your library as an explorable map. Auto-linking comes with Pro; manual linking is free for everyone.

Beyond document-document links, memories carry typed relations: derived_from, supports, contradicts, related_to, precedes, follows. Agents walk these trails to trace how knowledge was derived.

Plans gate capabilities, not accumulation

There are two plans — Free and Pro — plus a 14-day Pro trial on signup. The Free tier is generous on volume (1000 documents, 100 MB) but gates the power features: connectors (1), workspaces (2), API keys (1), and auto-linking (off). Pro lifts those caps and turns on the knowledge graph.

See Plans & Quotas for exact numbers.

The agent surface

Agents talk to Kiomon over two protocols:

1. MCP (Model Context Protocol) — the memory-native toolset: list_workspaces, get_workspace_context, search_memory, get_memory, explore_memory_graph, draft_memory, reflect_session, manage_memory. Used by Cursor, Claude Desktop, Windsurf, Cline, and any MCP client.
2. OpenAI-compatible API — drop-in replacement for OpenAI endpoints (chat/completions, embeddings) so n8n, LangChain, openai SDKs, and any agent tool can point at your Kiomon brain by swapping a base URL and API key.

Both authenticate with a Kiomon API key.