Walden3
A Matrix chatbot that remembers people, room by room, and keeps that memory on Walrus: encrypted, verifiable, and owned by the group. It runs on a local open-weight model, so no conversation is sent to a model provider.
2:43 · Download MP4
The problem
Groups like a class, a research team or a community talk over weeks and months. Chatbots in those rooms either forget what people told them, or they make it up. The ones that do remember keep the memory in a vendor's database, where the group can't verify it, move it or take it with them.
Same model, same room, same question. Real output, 8 October 2026.
What Walden3 does
Facts with provenance
Every fact records who said it, when, in which room and in which conversation. A person's memory holds only what they said about themselves.
Scoped to the room
Facts never leave the room they were learned in. The code enforces this. Every conversation is also summarised as an episode.
Checkpoints on Walrus Memory
When a session ends, the room's memory is encrypted on Walden's server and mirrored to Walrus Memory on mainnet. Each checkpoint links to the previous one in a hash chain.
Verify & load
!walden verify downloads a checkpoint from mainnet, decrypts it and checks the chain. !walden load carries memory into another room, even on another homeserver.
Consent-based archives
!walden archive preserves a whole room. Each person opts in. Episodes, memories and files are all encrypted with the archive password.
Local open model
IBM Granite 4.0 H-Small (32B MoE, 9B active) runs via Ollama on one GPU. Strict JSON schemas and gleaning make it a reliable memory writer.
Why it matters
With Walrus, the memory belongs to the group and not to the bot's operator. It is encrypted before it leaves the server, so the relayer sees only ciphertext. Anyone can check it on mainnet, it can move between rooms, and it can outlive the bot itself. Walden3 grew out of teaching and research at the Universität der Künste Berlin and the book project Walden 3 on education in times of AI. It was built for real groups, not for a demo.
In real use
- Running since 24 September 2026 in rooms on three Matrix homeservers
- 39 memories on Walrus mainnet under one Walrus Memory account (count from 9 October), across room checkpoints and the first archive (33 episodes)
- Memory moved between homeservers through Walrus alone: 22 facts loaded, decrypted and hash-verified
On-chain identifiers
| Walrus Memory agent ID (MemWalAccount) | 0x1979de15948d9fb7994ec9a78e313e7a454818050e2853142e81f13c73282094 |
| Package ID (Walrus Memory, mainnet relayer) | 0xe7c16fbea0560e7057e2bf7422feaa4fb313749fc69c9e9092fac7a33b81d7f5 |
| Dedicated wallet (account owner) | 0xbe54a8cf260e068dea6c4ce23e91c430af6cbd10791289e0c2d9e64b359473d5 |
| Example verified checkpoint blob | nV_tv-BFJ5iZzArog8oT2pQg3OWJ95PTu_fzu3NO9fk |
Try it
Join #walden3-demo:udk.ai or invite @walden:udk.ai into any Matrix room. Ask it something, tell it about yourself, then type !walden close and ask again.
Source code: github.com/hromi/walden3 · Whitepaper: Walrus memory graphs (PDF)