Graphlit

Graphlit is a cloud platform built by Unstruk Data Inc. that serves as the context layer for AI agents, providing organizational memory through structured entity extraction, relationship modeling, and continuous content ingestion. It is currently winding down — the pricing page notes that new signups are closed and existing accounts remain available during the transition. The platform is accessed via TypeScript, Python, and .NET SDKs, and exposes an MCP server for direct integration with tools like Cursor, VS Code, Claude Desktop, and ChatGPT.

What It Is

Graphlit positions itself as organizational knowledge infrastructure rather than a simple vector database or chat memory tool. Where vector databases store static embeddings and chat memory tools persist per-user conversations, Graphlit models entities, relationships, temporal state, and provenance across an organization's full content corpus. The platform ingests documents, audio, video, web pages, emails, Slack messages, Jira tickets, and more — then extracts structured knowledge that AI agents can query reliably. The result is a single API that covers ingestion, enrichment, storage, and retrieval without requiring developers to stitch together separate services.

How the Knowledge Model Works

Graphlit's core abstraction is the "Observable model" — a structured representation of entities (people, companies, decisions), events (changes, updates), ownership, temporal state, and relationships. According to the platform's documentation, this goes beyond cosine similarity search by resolving entities using LLM-powered matching rather than rule-based heuristics. Every answer is traceable to its source through full provenance tracking, which the platform claims dramatically reduces hallucinations compared to ungrounded retrieval.

Key modeling capabilities include:

• Entity extraction — people, companies, and relationships pulled automatically from ingested content
• Knowledge graph queries — graph traversal, entity co-occurrence, and relationship queries via queryGraph and queryFactsGraph methods
• Temporal state — how entities and facts evolve over time
• Provenance — cited sources traceable to origin for every retrieved answer

Integration Breadth and MCP Support

Graphlit connects to over 30 data sources through built-in connectors, including Google Drive, OneDrive, Dropbox, Box, SharePoint, Amazon S3, Azure Blob, Slack, Gmail, Outlook, Google Calendar, Microsoft Calendar, Notion, Confluence, GitHub, GitLab, Linear, Jira, Zendesk, Intercom, HubSpot, Attio, Reddit, Twitter/X, LinkedIn, Discord, Microsoft Teams, Zoom, Fireflies, Fathom, RSS feeds, and website crawling. The platform also ships an MCP server (npx -y graphlit-mcp-server) that connects Graphlit's knowledge layer directly to MCP-compatible clients. AI model support spans OpenAI (including GPT-5.4 via the Responses API), Anthropic Claude, Google Gemini, xAI, Deepseek, Groq, Mistral, Cohere, Cerebras, and AWS Bedrock — all with tool calling and streaming.

TypeScript SDK and Agent Framework

The TypeScript client SDK (MIT-licensed, available on GitHub) is the primary developer interface. Recent releases have added significant agent-oriented capabilities:

• v1.6.0 introduced a first-class Agent entity, the runAgent() autonomous harness with stuck detection, a Skills API for reusable agent skills, and an agent scratchpad for persistent state
• v1.5.0 added context management with token windowing, artifact collection, conversation queueing, and thinking/reasoning storage across turns
• Streaming Provider Resilience — streamAgent and runAgent now retry transient LLM provider errors (HTTP 500, 503, 429, network failures) with exponential backoff across all 10 supported streaming providers
• OpenAI Responses API — GPT-5.4 and eligible models are automatically routed through the Responses API for improved intelligence, lower latency, and higher cache hit rates

Current Status: Winding Down

The pricing page explicitly states: "Graphlit is winding down. New signups are closed; existing accounts remain available during the transition." A dedicated sunset details page is linked. Existing customers can continue to sign in and manage usage-based credits, and the support contact (questions@graphlit.com) handles account, transition, and data export questions. The founder's manifesto on the site describes Graphlit as the agent memory substrate, with a companion product called Zine (at zine.ai) serving as the human memory interface — both developed by Kirk Marple, Founder and CEO of Unstruk Data Inc.