How Tvix Works

How context flows into Tvix's two knowledge stores (Memory and Wiki), gets structured, and comes back to your agents.

Tvix transforms raw context into structured, retrievable knowledge that your agents can pull from across sessions. This page walks through the full lifecycle.

Architecture Overview

Tvix keeps two complementary knowledge stores, each optimized for a different shape of information. Agents can read and write both through MCP, and Chat in Dashboard can consult both when stored context helps.

Context from either store flows through three stages: ingest → process → retrieve. The rest of this page walks through each stage for both stores.

1. Context to Tvix

How context enters each store.

Memory

Memory receives context from imports, agents, Chat in Dashboard, and integrations.



  

Chat History Import (bootstrap)

    Already have months of conversations in ChatGPT, Claude, or Gemini? Export your chat history and upload it in [Sources](https://app.membase.so/sources) under the **Chat History** section. The entire archive goes through the same digesting pipeline as live conversations.

    

Live agent conversation

    You talk to your agent normally. During the conversation the agent picks up on preferences, decisions, project details, and other durable context, and calls `add_memory` via MCP.

    ```text Example theme={null}
    You: "Let's use Zustand instead of Redux for this project.
          Also, I prefer functional components over class components."

    → Agent calls add_memory with this context
    ```
  


  

Chat in Dashboard

    While talking directly to your knowledge base in the dashboard, Chat can save durable personal context as memory when you share something worth keeping.
  


  

App integration sync

    Connected sources (Gmail, Google Calendar, Slack) sync new data automatically in the background. Each message, event, or email becomes an episode.
  


  

Tvix receives the context

    Tvix accepts the incoming context for processing so the agent or sync can keep moving.
  

Wiki

Wiki receives documents directly from you, Chat in Dashboard, or agent tool calls.



  

Wiki import (bootstrap)

    Already have a markdown knowledge base? Use **Wiki > Add > Import files** to import Notion exports, Obsidian vaults, or Markdown files into a target Project or Basic. `[[wikilinks]]` are preserved where available, and documents upload through the background import flow. See [Importing Documents](/features/wiki#importing-documents).
  


  

Write a document in the dashboard

    Open the Wiki tab, click **Add > Write document**, choose a Project or Basic, and write in markdown. Manual save and unsaved-changes warnings keep your content safe.
  


  

Notion live sync

    Connect Notion from Sources, choose the Notion pages available to Tvix, then choose a target Wiki Project. Synced pages become source-backed Wiki documents. See [App Integrations](/connectors/apps#notion-live-sync).
  


  

Agent calls add_wiki

    Your connected agent can call `add_wiki` when you share factual, reference-style knowledge worth keeping. Example: "Write up our deployment rollback checklist" results in a new wiki document instead of a memory.
  


  

Supported plugin captures transcript

    Supported agent plugins can preserve user/assistant conversation transcripts as original source material in Wiki, separate from extracted personal Memory.
  


  

Chat creates a document

    Chat in Dashboard can also create wiki documents when you ask it to save factual, reference-style material.
  


  

Tvix receives the document

    The Wiki document is saved directly or prepared for background processing, depending on whether it came from a direct write, sync, or import.
  

2. Digesting and Structure

How raw input becomes structured, searchable knowledge.

Memory

Every memory (from agents or integrations) goes through the same pipeline.



  

Episode creation

    The raw input is saved as an **episode**, a snapshot of a conversation or data event. Episodes are the building blocks of memory.
  


  

Entity extraction

    Tvix identifies key entities from the episode: people, projects, tools, preferences, decisions, dates, and other meaningful concepts.

    ```text Example entities from a conversation theme={null}
    "Let's use Zustand instead of Redux"
    → Entities: Zustand, Redux, state management decision
    ```
  


  

Graph construction

    Extracted entities are added to your knowledge graph. New entities connect to existing ones when they overlap, so "Zustand" mentioned in two different conversations becomes one entity with two linked episodes.
  


  

Deduplication and merging

    If the same fact appears in multiple episodes, Tvix merges them. When new information contradicts an existing memory, the latest data takes priority.
  




The result is a continuously growing knowledge graph where entities, relationships, and episodes are all interconnected.

Wiki

Wiki documents are stored as markdown plus structural metadata for fast retrieval.



  

Document storage

    The document's title and markdown content are saved to the Wiki store, filed into a Project or Basic.
  


  

Wikilink parsing

    Tvix parses `[[wikilinks]]` in the content and maintains a bidirectional link graph. This is what powers backlinks, the force-directed graph view, and instant `[[` autocomplete in the editor.
  


  

Search indexing

    Each document is indexed for **hybrid search**: a full-text index (BM25-style) for keyword queries and a semantic embedding for meaning-based queries. Both are fused with Reciprocal Rank Fusion (RRF) at query time.
  


  

Source provenance and review

    Imported and synced documents keep source provenance when available. When a source update or generated change needs review, Tvix can route it through Inbox Review so you can inspect, edit, accept, reject, or undo the change.
  

3. Tvix to Your Agent

How context flows back when an agent or Chat needs it.

Memory

Agent calls search_memory

    The agent sends a query describing what personal context it needs. This happens automatically when prior context would improve the response.

    ```text Example theme={null}
    You: "Set up a new component for the settings page."

    → Agent calls search_memory: "project tech stack, component preferences"
    ```
  


  

Tvix searches the knowledge graph

    The query is matched against your graph using semantic search. Relevant episodes and their connected entities are retrieved.

    Relevant results might include:

    * "Uses Next.js with TypeScript"
    * "Prefers functional components"
    * "State management: Zustand"
    * "Styling: Tailwind CSS"
  


  

Ranked results returned

    Results are scored by relevance and returned as episode-centric bundles. Only the most useful context is included, keeping the agent's context window clean.
  


  

Agent responds with full context

    The agent generates a response grounded in your actual preferences and project details, without you having to repeat any of it.
  




Here's a real example: Claude retrieving a git workflow from Tvix during a conversation.

<video autoPlay muted loop playsInline>
    <source src="https://mintcdn.com/aristo_2/SyC5fNZLZj3t6Psb/images/core-concepts/tvix-to-agent.mp4?fit=max&auto=format&n=SyC5fNZLZj3t6Psb&q=85&s=a42b0c5d241b25d307141230b9c3c2f5" type="video/mp4" data-path="images/core-concepts/tvix-to-agent.mp4" />
  </video>

Wiki

Agent calls search_wiki

    The agent sends a query describing what factual knowledge it needs.

    ```text Example theme={null}
    You: "Remind me how our auth middleware handles expired tokens."

    → Agent calls search_wiki: "auth middleware expired token"
    ```
  


  

Tvix runs hybrid search

    The query is matched with both full-text keyword search and semantic similarity. The two rankings are fused with Reciprocal Rank Fusion, so you get solid results whether the query is literal or conceptual.
  


  

Full document bodies returned

    Unlike memory episodes, wiki results include the **full document body**, so the agent has enough context to answer directly instead of juggling fragments.
  


  

Agent responds with full context

    The agent grounds its answer in the retrieved documents and cites them if your prompt encourages citations.
  

Tip: Chat in Dashboard can use both search_memory and search_wiki when stored context could help, then combine the results into a single answer with citations. Your agents should do the same when the user's question could benefit from either store.

This entire cycle (ingest → structure → retrieve) runs continuously as you use Tvix. The more you interact, the richer both stores become, and the smarter your agents get.

Next Steps

Attached vs Universal

Understand why shared memory matters.

Tvix MCP

Learn about the MCP tools available to your agents.

Memory

Dive into memory exploration and management.

Knowledge Wiki

Dive into Wiki documents, Projects, imports, and Inbox Review.