Locus

Verifiable AI memory on Solana — built on ARMS.

Solana agents are stateless. The Agent Payments Protocol launched this quarter and agents are trading, voting, and coordinating on-chain — but none of them have persistent, verifiable, queryable memory. Locus is the Solana protocol that turns ARMS, a spatial memory fabric in Rust, into a payable memory network: memory roots committed on-chain, pay-per-query economics, and a cryptographic retrieval attestation for every AI-driven decision.

This repo is the hackathon submission for the Solana Frontier Hackathon (Colosseum).

---

Live on devnet

Proof artifacts you can click right now:

• Program ID: C6AJ43ZpzPLtmcwDS1FQP7cQXtWHNwsLty5ijdLTxzmK
• Program deploy tx: 2V2Yrot…d2n
• initialize_agent tx: 2FPgj7p…RLZ
• commit_memory tx (root bd8e1419…5f16): 3EJotzn…AywF
• attest_retrieval tx: 4UxWaB1…j8iX
• RetrievalAttestation PDA: 9z8SwfDQRu2Nt1wbtQooLitpb3WQXge8PzxdyYebQLVt

What the attestation records (all on-chain, all verifiable):

Field	Value
agent	PDA [b"agent", owner]
version	1
memory_root	bd8e1419cf6699efd55d6a573f9c6a2f942a51c3c593f22e19fddf5caf8f5f16
query_hash	d5d98ebe29342856fb0f3e6d6412f219122e825b0f8b6164c85b72234587946f
result_hash	4553cb46fa3ec8d8aed75376d154667094cebce6e5c1ce2a3b76ac495cdd11b6
nonce	16070694091605954182

Behavioral similarity result (query wallet was Bubblegum-heavy → neighbors are also Bubblegum-heavy):

=== top 5 behaviorally similar wallets to 1tjoQzWg21ceMAhyLVs4s2Zc1Pm27mLTcJ9EfEFu7km ===
   1. 1tjoQzWg21ceMAhyLVs4s2Zc1Pm27mLTcJ9EfEFu7km   dist=0.0000
   2. AGkGWK1R669KDT4FCqgDgK7PgahGJPjD4J9xmVjuL9kn  dist=0.8792   bubblegum + compressed NFT
   3. 5oNhn2ESFHgg52xiEKBt93cEEnbsoUoekcxXsWpLbHUJ  dist=0.9779   bubblegum + transfers
   4. 5SybwTvGno6SBApbGinXyLqhnPzs47aaBG4PShyPT8ey  dist=1.1390   high-volume transfers
   5. 9xSi6onCqhLuXTgodd5wB7EXv2WwuYPZyZwDyruGKL6v  dist=1.1918   compressed NFT mint heavy

---

Pitch video — generated end-to-end with local AI

The 3:00 pitch video lives at demo/locus-pitch-final.mp4. Every frame and every word was produced locally on an RTX 5090 — no cloud TTS, no stock footage, no human voiceover. Real Solscan footage for the on-chain proof, real ARMS paper figures for the visuals, real Qwen3-TTS narration for the voice. We ship AI with AI.

Visual recording	Playwright (chromium, 1920×1080) — drives a self-paced HTML deck and live explorer.solana.com navigation
Voice narration	Qwen3-TTS-12Hz-1.7B-CustomVoice (1.7B params, local, ~4.3 GB) with per-section emotion instruct
Figures	Real ARMS paper figures from Lumi-node/ARMS (DOI on Zenodo)
On-chain proof	Real https://explorer.solana.com/ pages — every tx + account decoded live during recording
Audio post	ffmpeg loudnorm I=-16 LUFS for broadcast-level output

Reproducible from this repo:

# 1. record the 6-section pitch deck at 1080p (3:02 webm)
npx playwright install chromium
npx playwright test scripts/demo-solscan.spec.ts

# 2. synthesize TTS narration (edge-tts, neural voice, auto-paced per segment)
pip install --user edge-tts
python3 scripts/generate-narration.py

# 3. mux narration onto the recording, output a YouTube-ready MP4
./scripts/mux-demo.sh

Each section's narration text + timing window lives in demo/narration.json — edit + re-run to retune voice, pacing, or copy.

30-second demo

# 1) one-time setup
./scripts/airdrop.sh                       # creates dev keypair + airdrops devnet SOL
./scripts/deploy-devnet.sh                 # builds + deploys the Anchor program

# 2) run the demo
./scripts/seed-demo.sh                     # spins up ARMS service + ingests wallets
cargo run --release -p locus-demo -- similar --wallet <pubkey> --k 10

That last command finds the 10 most behaviorally similar Solana wallets to the one you pass, calls attest_retrieval on devnet, and prints the on-chain attestation signature.

---

Architecture

                +-------------------+
                |    locus-demo     |       CLI: ingest / similar / show-attestation
                +---------+---------+
                          |
                          v
+-------------+   +-------+--------+        Rust SDK: agent ops + attested queries
|   wallet    |-->|  locus-client  |
+-------------+   +-------+--------+
                          |
              +-----------+-----------+
              |                       |
              v                       v
   +------------------+      +------------------+
   |  arms-service    |      |  Locus program   |     Solana devnet
   |  (Axum, wraps    |      |  (Anchor)        |
   |   arms-core)     |      |                  |
   +--------+---------+      +------------------+
            ^                          ^
            |                          |
            |   +------------------+   |
            +---+  locus-relayer   +---+    poll ARMS, merkle, commit_memory
                +------------------+

• programs/locus/ — Anchor program: AgentMemory + RetrievalAttestation accounts, initialize_agent / commit_memory / attest_retrieval / update_read_fee / close_agent instructions.
• locus-client/ — Rust SDK. Agents/devs talk to Locus + ARMS through one type (LocusClient).
• locus-relayer/ — long-running process that recomputes the ARMS state's Merkle root and commits it on a poll interval.
• locus-demo/ — end-to-end CLI: pull Helius tx history, build behavioral feature vectors, push to ARMS, run attested similarity queries.
• arms-service/ — thin Axum HTTP wrapper around the arms-core crate (ARMS itself stays untouched).

---

Who this is for

Locus is the audit layer for any AI-driven on-chain action. Four concrete users:

Trading agents. When an autonomous agent makes a $10K on-chain decision, the attest_retrieval PDA records exactly which memories shaped it — query hash, result hash, memory root, version. Reconstruct decision lineage after the fact. Insurance, post-mortem, compliance.

Governance agents. DAO votes are increasingly delegated to AI. Locus lets the delegated agent prove which historical proposals, votes, and discussions informed each new vote. Audit a governance agent the way you'd audit a human delegate.

DePIN coordinators. Multi-agent fleets sharing operational memory across nodes need a verifiable common substrate. The on-chain Merkle root is the consistency primitive; ARMS is the geometry beneath it.

Compliance auditors and regulators. Reading attestation PDAs is free — anyone with a Solana RPC can reconstruct the decision graph of an agent's operating history. Audit-as-a-service without trusted intermediaries.

The same primitive — memory commitments + paid attested retrievals — powers all four.

---

Build from source

Requirements (./scripts/install-toolchain.sh automates these on Linux — review first):

• Rust 1.79+ (rustup show)
• Solana CLI 1.18+ (solana --version)
• Anchor 0.30.1 (anchor --version)
• Yarn or npm (for the TS tests under tests/)

# Build the Solana program (BPF/SBF)
anchor build

# Build all Rust crates (host)
cargo build --release

The Locus program declares its ID in programs/locus/src/lib.rs via declare_id!(...). After the first anchor deploy, copy the printed program ID into:

1. declare_id!(...) in programs/locus/src/lib.rs
2. The [programs.devnet] entry in Anchor.toml
3. LOCUS_PROGRAM_ID in your .env

Then rebuild and redeploy once so the on-chain ID matches the embedded ID.

---

Run from scratch

# 1. start the ARMS service (in-memory; dim 64)
cargo run --release -p arms-service -- --dim 64

# 2. start the relayer (in another shell)
export LOCUS_KEYPAIR=$HOME/.config/solana/locus-dev.json
export LOCUS_PROGRAM_ID=<your deployed program id>
cargo run --release -p locus-relayer -- \
  --agent <owner pubkey> --arms http://localhost:8080 --interval 30

# 3. ingest wallets + commit root
cargo run --release -p locus-demo -- ingest \
  --wallets data/wallets.json --limit 10000 --dim 64

# 4. attested similarity query
cargo run --release -p locus-demo -- similar \
  --wallet <pubkey> --k 10

---

Background research

Locus is not a hackathon-spawned concept — it productizes pre-existing AI memory research onto Solana.

• ARMS — Attention Reasoning Memory Store. Spatial memory fabric in Rust. Position IS relationship. → github.com/Lumi-node/ARMS · Zenodo DOI
• HAT — Hierarchical Attention Tree retrieval algorithm. 100% recall, 70× faster build than HNSW. → github.com/Lumi-node/HAT
• infinite-context — productized HAT variant: 11M+ tokens, 0.51 ms latency, 100% recall. → github.com/Lumi-node/infinite-context
• HuggingFace models — AutomateCapture/HAT · AutomateCapture/ARMS
• Author — Andrew Young, independent AI researcher. ORCID: 0009-0002-3685-4289

---

Roadmap

• Mainnet deployment with a token-incentivized DePIN coordinator for ARMS nodes
• ZK proofs of retrieval (replace the SHA-256 result hash with a SNARK over the actual neighbor set)
• First-class integrations with Solana agent frameworks (Agent Payments Protocol, Jupiter agents, governance agents)
• Onboarding the first 10 agent teams via the Colosseum accelerator

---

Team

Andrew Young — independent AI researcher (Lumi Node / Automate Capture LLC). andrew@automate-capture.com

License

MIT — see LICENSE.