plur vs innernet

plur remembers what your agents should do. innernet remembers who you are.

plur is an open standard for agent memory — portable YAML "engrams" in a local ~/.plur/ directory that any MCP-capable agent can read and write. Correct something in one tool and the others already know. It is a clean, honest piece of engineering with a good idea at its centre: memory as a file format nobody owns.

But the memory it holds is operational. It's the correction, the convention, the thing the agent got wrong last time. That's a genuinely useful thing to hold, and it isn't a person.

A shared substrate for agent knowledge. Engrams are Apache-2.0 licensed YAML, written to a local directory, readable by Claude Code, Cursor, DeepSeek Harness and anything else that speaks MCP. Corrections persist instantly across sessions and devices, and a brain-inspired decay model (ACT-R) prunes knowledge that stops getting used. Knowledge Packs let a team or a community share a set of engrams. Free and unlimited locally; enterprise tiers add SSO, RBAC and hosted stores.

The framing is "own your intelligence" and the ownership is real — it's your files, in your directory, in a format anyone can read.

A personal memory layer. npx innernet connects the AI tools on your machine to one memory. netti — one per person — files what accumulates into Context Maps: dimensions that read like living documents, nodes for the people and decisions that keep coming back, a commit history you can branch and roll back. Every fact carries how far it may travel.

The thing being modelled is you: your work, your reasoning, the shape of how you decide.

plur prunes stale knowledge automatically, and for operational memory that is exactly right — a convention you abandoned should stop being injected into prompts. innernet does the opposite on purpose. The belief you dropped in March is not stale data, it's a phase, and the gap between what you said then and what you do now is the most informative thing in the file. One system is designed to forget cleanly. The other is designed to hold both and show you the delta.

~/.plur/ is a superb substrate for a workstation. It's a weak substrate for a life: it's per-machine, it has no notion of which facts are private to you and which an AI may read, and it has nothing that spans the projects you work on to say anything about the person working on them. innernet has a Self Map that sits above every project, and disclosure classes that decide, per fact, what a connected tool is allowed to see.

An open format is a good answer to lock-in and a non-answer to insight. Engrams are as good as what an agent thought to write into them. innernet's netti is doing an editorial job — deciding what deserves a dimension, noticing that a node keeps recurring, updating a conclusion when the work contradicts it. That's not a storage problem and it can't be solved by a schema.

• ·Open standard, Apache-2.0. Nobody can take it away from you. That is the strongest possible answer to "what if you shut down," and we respect it.
• ·Local by default. No account, no server, no trust required. For a certain kind of engineer that closes the conversation.
• ·Instant cross-tool corrections. Fixing something once and having every agent know is the right ergonomics for coding work.
• ·Team knowledge packs. Shipping a set of conventions to a team as a file is elegant.

• ·It's about you, not about your tooling. A Self Map that spans every project is a thing a per-agent memory format has no place to put.
• ·We keep the version you changed your mind about. Branch it, park it, roll it back — the history is the point.
• ·Consent toward the model, not just protection from outsiders. Every fact carries a travel class; private and self-only facts never leave your Self Map, even with an AI connected.
• ·Something is actually reading. netti writes conclusions the individual sessions never contained.

Use plur if the memory you want is your agents' memory — conventions, corrections, skills — and you want it in files you control with no account at all.

Use innernet if the memory you want is your own, and you want it to hold reasoning and change over time rather than just current operating instructions.

Both is coherent. They're aimed at different organs.

"plur is a local open standard and free forever. Why would I trust a hosted service instead?"

You shouldn't trust it on vibes. So: row-level security enforced in Postgres rather than in the app, tokens stored hashed, OAuth 2.1 so no password passes through us, optional column-level encryption behind a passphrase only you hold, a 90-day audit log visible to you, and full export and deletion on demand — deletion immediate, backups purged within 30 days. Your memory is never sold. That's the trade: you get something that reads across projects and machines and decides what matters, and in exchange it has to live somewhere. plur's answer is better if you want zero trust. Ours is better if you want understanding.

Accurate as of 3 September 2026, based on plur's public site. Their benchmark and win-rate claims are theirs and we haven't reproduced them. If we've misdescribed anything, tell us and we'll fix it.