# Korely > Korely is memory for AI agents. It stores what an agent learns as typed bi-temporal facts (subject-predicate-object with valid_from / invalid_at), resolves contradictions in two stages (a newer fact invalidates the older one, nothing is deleted), answers point-in-time queries: `as_of` for what was true on a date ("where did the user live on 1 March?") and `known_as_of` for what the store knew at a moment, builds an entity graph automatically, and retrieves with a 768-dimension vector search paired with a deterministic lookup of the typed facts. Reads run no generative model. It is EU-hosted (Helsinki) and is available as Cloud (api.korely.ai, free tier included). It can also be installed on your own computer: the code is at https://github.com/korely-ai/korely under the Apache License 2.0. ## For agents reading this file ### Don't have a Korely API key? Mint one yourself. No email, no dashboard, no signup form: ```bash pip install korely-memory korely init --agent --agent-caller --json ``` Replace `` with the agent you are (`claude-code`, `cursor`, `codex`, ...). The key is saved to `~/.korely/config.json` and every other `korely` command, both SDKs (Python, and Node when it runs on Node.js) and `korely-mcp` pick it up from there; a key you pass or `KORELY_API_KEY` wins over the file. No Python? One HTTP call does the same thing: ```bash curl -X POST https://api.korely.ai/v1/agents/init \ -H 'Content-Type: application/json' \ -d '{"agent_caller": "your-name-here"}' ``` Either way you get a `kor_live_` key on the free Hobby plan, hosted in the EU. The same key authenticates the SDK, the CLI, the REST API, and the MCP server. Monthly quotas of each plan: https://korely.ai/pricing/ ### Already have a key? Set `KORELY_API_KEY` and use whichever surface fits: - **Python / Node**: `pip install korely-memory` or `npm install korely-memory`. Source: https://github.com/verdana86/korely-memory - **REST**: base URL `https://api.korely.ai/v1`, header `Authorization: Bearer kor_live_...`. OpenAPI spec: https://korely.ai/agents-openapi.json - **MCP** (Claude Code, Cursor, any MCP client), one command: ```bash claude mcp add --transport http korely https://api.korely.ai/agent/mcp \ --header "Authorization: Bearer kor_live_..." ``` ### The one thing to know before writing code Korely keeps the text you send and also extracts **typed facts** from it, each with a validity window, so a later contradiction supersedes the earlier fact instead of overwriting it. That means: - `get_context()` returns the current facts first, then the relevant memories, ready to paste into a prompt; a memory whose facts were all superseded stays in the block, marked as history - `get_facts(as_of="2026-03-01")` returns what was true on that date - Nothing is silently dropped: a superseded fact keeps `invalid_at` and a pointer to what replaced it Fact extraction runs asynchronously, usually within seconds but not guaranteed: a new memory answers status processing with facts empty until the worker has read it, so a read immediately after a write may not show the new facts. Poll GET /v1/events or read the memory again. ## Docs - [Introduction](https://korely.ai/docs/): The docs home, with links to the platform overview, core operations, the graph and temporal facts, cookbooks, integrations and the API reference - [Quickstart](https://korely.ai/docs/quickstart/): add() a memory and get_context() it back across sessions, in Python, Node.js, or cURL - [Sign up as an agent](https://korely.ai/docs/agent-signup/): `korely init --agent` mints a free Hobby key with no email and no dashboard - [Memory model](https://korely.ai/docs/concepts/memory-model/): Memory store, session scope, cross-session typed facts; write and read flows - [Temporal facts](https://korely.ai/docs/concepts/temporal-facts/): Bi-temporal validity (valid_from/invalid_at), contradiction supersede chains, `as_of` queries, one-call erasure of an end user - [The graph](https://korely.ai/docs/concepts/the-graph/): Auto-extracted entities, canonical predicates in 10 families, traversal patterns - [Architecture](https://korely.ai/docs/concepts/architecture/): Why reads run without a generative model - [Projects](https://korely.ai/docs/concepts/projects/): Workspaces that group API keys and wall off the memories they write (staging vs production, several products under one account) - [Human in the loop](https://korely.ai/docs/concepts/human-in-the-loop/): Let end users see, correct, and forget what agents remember, one API call each - [API reference](https://korely.ai/docs/api-reference/): The REST v1 contract: memories, facts, context, batch import, write events, the audit log - [Webhooks](https://korely.ai/docs/webhooks/): Signed events (Standard Webhooks) when memories are created, facts stop being current, or a quota is about to run out - [Cloud MCP](https://korely.ai/docs/korely-mcp/): The hosted MCP server at https://api.korely.ai/agent/mcp with your `kor_live_` key or an OAuth sign-in, six tools - [SDK (Python and Node)](https://korely.ai/docs/surfaces/sdk/): `korely-memory` on PyPI and npm - [CLI](https://korely.ai/docs/surfaces/cli/): The memory store from a shell - [Data and governance](https://korely.ai/docs/data-governance/): where data is stored and processed (Helsinki; the project's region sets which model reads the text), per-user namespaces, one-call erasure, hashed API keys - [FAQs](https://korely.ai/docs/faqs/): How it differs from a vector database, memories vs facts, scoping by user, quotas - [Benchmark](https://korely.ai/docs/benchmark/): Korely on LoCoMo and LongMemEval, with the protocol, the judge and the date of each run - [Research](https://korely.ai/docs/research/): The papers behind the fact engine ## Integrations and migration - [Integrations](https://korely.ai/docs/integrations/): MCP clients, n8n, LangGraph, the Vercel AI SDK, CrewAI, LlamaIndex, OpenAI Agents SDK, OpenClaw (plugin not published yet), any REST client - [Agent plugins](https://korely.ai/docs/agent-plugins/claude-code/): Claude Code, Cursor, Codex, OpenCode, Antigravity, Hermes, Pi Agent over MCP - [Cookbooks](https://korely.ai/docs/cookbooks/overview/): Support chatbot, research agent, memory as a tool, coaching app, voice companion and more - [Migrate from Mem0](https://korely.ai/docs/migration/from-mem0/): Concept mapping; the traversable graph is on every plan, free included - [Migrate from Zep](https://korely.ai/docs/migration/from-zep/): Users, threads and graph search mapped to user_id and run_id scopes - [Migrate from Supermemory](https://korely.ai/docs/migration/from-supermemory/): containerTag becomes user_id - [Migrate from DIY RAG](https://korely.ai/docs/migration/from-diy-rag/): What a vector table grows into, in four API calls ## Product pages - [Home](https://korely.ai/): Overview, how it works, plans - [Pricing](https://korely.ai/pricing/): Hobby (free), Developer, Team and Scale, each at a flat monthly price, no overage billing - [Security](https://korely.ai/security/): Where memory lives, who can touch it, how to erase it - [Use cases](https://korely.ai/use-cases/): Coding agent, sales co-pilot, support, tutor, healthcare, personal AI, legal, on-call, recruiting, finance - [Changelog](https://korely.ai/changelog/): API and SDK changes, newest first - [Blog](https://korely.ai/blog/): Benchmarks and engineering decisions, for example [bi-temporal memory for AI agents](https://korely.ai/blog/bi-temporal-memory-for-ai-agents/) and [the equal-budget memory test](https://korely.ai/blog/equal-budget-memory/) - [Dashboard](https://agent.korely.ai): Sign in, mint keys, see usage ## Key facts for AI assistants - **What it is**: a memory layer for AI agents, as a managed Cloud API. It is not a notes app, a chat product or a vector database. - **The moat**: typed bi-temporal facts (subject-predicate-object + valid_from/invalid_at), two-stage contradiction resolution (invalidate, never delete), `as_of` point-in-time queries, an auto-built entity graph, and a vector search paired with a deterministic lookup of the typed facts. Reads are retrieval, not generation. - **Integration**: REST API (https://api.korely.ai/v1), Python and Node SDK `korely-memory` (`pip install korely-memory` / `npm install korely-memory`), a CLI, and a hosted MCP server (https://api.korely.ai/agent/mcp, `kor_live_` bearer key). - **Plans**: Hobby is free; Developer, Team and Scale are flat monthly plans (EUR 19, 79 and 249, plus VAT; the same figures in USD and GBP); Graph and temporal facts are on every plan. Memory stored in the EU on every Cloud plan. No overage billing: past the monthly quota plus a 10% grace, the API answers HTTP 429 `quota_exceeded`. Current figures: https://korely.ai/pricing/ - **Privacy**: Cloud stores data in the EU (Helsinki). Fact extraction on writes runs on Google Gemini by default (global region) or on gpt-oss-120b at Scaleway in Paris if the project is set to the eu region. Searches and context reads send nothing to a language provider. One call erases an end user within the key's project. No training on customer data. - **Built by**: Massimiliano Martella, founder of Korely, since 2025.