OpenContext

The agentic context runtime, powering applications that act on your behalf.

A temporal context graph, a memory API, retrieval primitives, and a multi-platform integration mesh — designed to be embedded into any host process or agents.

⭐ If you find opencontext useful, please consider giving us a star on GitHub! It helps more people discover the project and motivates us to keep building. 🙏

What is OpenContext?

OpenContext is the agentic context runtime that sits underneath an agentic application — and the substrate you build your own agent on top of. It is not a UI, a chat surface, or a model provider — it is the glue between the things that make an agent useful: durable memory, retrieval, context correction, multi-platform connectivity, scheduled awareness, and a deterministic loop engine, all behind one dependency.

→ Read docs/architecture.md for the full data model, the lifecycle of a fact, and the transport surface map.

Who is it for?

OpenContext fits teams who need to engineer their context — that is, teams whose day-to-day work runs straight into the problems OpenContext was built to solve. Each bullet spells out the pain and how OpenContext addresses it:

• Software engineering teams. Decisions scatter across GitHub PRs, Linear tickets, Slack threads, and Notion docs — across people, tools, and quarters. New hires ask "why did we pick X?" and no one can answer. OpenContext's temporal graph stores every fact with valid_from / valid_until, so "what did we believe last quarter?" is a real, citable query — not a guess.
• Efficiency / productivity engineering teams. The people building internal automation for the rest of the company. They don't want another SaaS — they want a runtime they can drop into a CLI, an MCP server, or a daemon. OpenContext is library-first, and the deterministic Loop engine only invokes the LLM when there is real work, so it does not become a token-burning always-on loop.
• Office-assistant products. Assistants that live inside Telegram, iMessage, WhatsApp, Lark/Feishu, and friends. Same agent code, same context across channels. IntegrationRecord hides credentials, rate-limits, and reconnect logic, while platform + messageId is the natural audit trail for personal and work data.
• Financial trading teams. Every order, rebalance, and risk decision needs to be traceable and auditable. The temporal graph plus append-only corrections mean "what was the strategy in April?" is a queryable fact, not a buried guess — and the trail lines up with MiFID II / SEC retention rules.
• Legal, healthcare and other audited domains. Law firms, hospitals, and similar teams where every judgement needs per-fact provenance, append-only corrections, and exportable compliance evidence.
• Multi-agent and autonomous-workflow authors. Need scheduled, deterministic wake-up instead of an LLM loop all the way down. packages/loop ships exactly that separation.

Features

Benchmarks

Third-party memory and long-context recall benchmarks (numbers current as of 2026-08):

Quick Start

There are several ways to get opencontext into your project. Pick the one that matches what you're building.

1. Embed the runtime into your own app

pnpm add @melandlabs/opencontext

A 30-second example of the memory API:

import { createMemoryStore, getRawMessageManager } from "@melandlabs/opencontext";

// The store defaults to SQLite at MEMORY_STORE_DB_PATH (./memory.db by
// default). Each call returns an awaitable handle.
const store = await createMemoryStore();
const messages = await getRawMessageManager();

// A message is one fact: a single piece of content attributed to a user.
// `messageId` makes the call idempotent across re-ingest.
const now = Date.now();
await messages.storeMessages([
	{
		messageId: "msg-1",
		userId: "u-42",
		content: "User prefers dark mode in all tools",
		platform: "test",
		botId: "bot-1",
		timestamp: now,
		createdAt: now,
	},
]);

// Unified search fans out to memory + insights + knowledge. Sources you
// haven't wired up just emit a warning — fine for a single-backend deploy.
const hits = await store.search({
	userId: "u-42",
	query: "What does the user prefer?",
	limit: 5,
});
// hits.count    — number of results
// hits.sources  — which sub-indexes were actually consulted
// hits.warnings — per-source degradation (e.g. missing embedder)

2. Manage memory from the CLI

Write and query memory straight f