A memory graph that knows when it's been wrong before.

Memagen™ is the open-core agent platform with a multi-domain compound-graph memory. It tracks what your agents know, when they learned it, and which facts have since been invalidated. Bring your own LLM. Run it locally. Keep your data.

Magnification level

1 6

00:00 00:30

The substrate visualization requires JavaScript. Read the substrate-as-live-visualization paper for the same content in prose form.

Loading scenario… Captures shown are simulated for demonstration; the recorder + anonymizer infrastructure is ready for live runs.

00:00 / 00:00

Memory for agents.

A graph-native substrate where what an agent learns becomes part of what it is.

How it works

Three steps. No SDK ceremony.

Install

$ pip install memagen-lite

One pip install. No Docker, no daemon, no account.

Configure your LLM provider

$ export ANTHROPIC_API_KEY=...   # or OPENAI_API_KEY, GEMINI_API_KEY,
                                # OPENROUTER_API_KEY, NIM_API_KEY

Bring your own key. Memagen™ never proxies your tokens, never stores your prompts. The model you pay for is the model you talk to.

Hand it a goal

>>> from memagen import Memagen
>>> m = Memagen()
>>> m.run("draft a release note for v0.4 from the changelog")

Memagen™ remembers what it tries, what works, what doesn't. Next time you ask, it starts from what it already learned, not from zero.

Most agent platforms forget. Memagen™ remembers.

Memagen™ has a compound graph that tracks every entity the agent encounters — user preferences, goals, capabilities, recipes, tools, the relationships between them — and the moment each fact entered the graph.

When a fact gets contradicted, the graph keeps the old fact, marks the moment it changed, and records what replaced it. Nothing is silently overwritten. Every belief is dated, every belief is revisable, every revision is auditable.

That is what gives Memagen™ continuity. Not a longer context window, not a vector store retrieving by similarity. A graph of things the agent actually learned, the order it learned them, and which of them are still true.

Read the design note →

One excerpt. Real graphs run thousands of nodes; the structure is the same.

One product page for humans. One for the agents who index for them.

Memagen™ has its own product page. We also have one your agents can read. Visit /agent for a structured-data version of this site, or /llms.txt for the LLM-friendly variant.

We're not the first company to do this — Lindy did it defensively, after seeing crawler traffic outpace human traffic. We're doing it because our customer is a developer who builds with agents, and agents are how their developers find new tools. If your agent can't read our page, we are invisible to the people we built this for.

/agent → Structured JSON-LD: capabilities, pricing, install, changelog, EULA

/llms.txt → Plain-prose summary for LLM context windows. RFC-style.

/llms-full.txt → Long-form: every doc page concatenated, freshness-stamped per file.

Why I built Memagen™.

I built Memagen™ because every agent platform I tried forgot what we'd already learned together. The model got smarter. My interactions with it didn't compound. Each new conversation re-met me, re-asked me what I cared about, re-discovered the same dead ends I'd already walked into a week earlier.

That bothered me enough to spend a year on it. Memagen™ has a compound-graph memory — multi-domain, with explicit fact invalidation, with a capability-grant audit trail, with the order of events preserved. When the agent learns something today, the agent you talk to next month starts from there.

It's open-core. The harness is on GitHub; if you want to read it, fork it, file an issue, that's the point. The graph engine is paid because the graph engine is what took the year. I would rather be honest about that than charge for the harness and pretend the graph is a feature.

Two things I won't pretend. One: Memagen™ isn't magic. It is a graph, a scheduler, and a careful set of rules about what to remember and when to mark something as no longer true. If the graph is wrong, the agent acts wrong. Two: this is early. The Connected tier is not shipping yet. The local tier is. I'd rather you start there, run it against your own work, and tell me what breaks.

If something here reads off, mail me. I read it.