MnemoPay Mobile SDK

On-device persistent memory (encrypted SQLite + sqlite-vec), agent-to-agent payments, and spatial proofs. TypeScript / Node 20+.

Development

npm ci
npm run lint    # tsc --noEmit
npm test        # unit tests (excludes tests/benchmarks/)
npm run build   # emits dist/

Crypto keys and migration

MnemoPay.create() wires NodeCrypto with:

• encryptionKey — AES-GCM; defaults to SHA256("mnemopay:" + agentId) when omitted.
• hmacKey — memory integrity HMAC; defaults to SHA256("mnemopay:mac:" + agentId).
• signingKey — Ed25519 seed; defaults to SHA256("mnemopay:sign:" + agentId).

Older builds only fixed the encryption key and drew random HMAC/signing material per process. That broke cross-device sync and manifest signatures. If you open an existing database after upgrading:

• Same device, same code: keys are now deterministic per agentId, so behavior is stable.
• Existing rows written under random HMAC keys may fail integrity verification on recall unless you still have the old keys. For production, set encryptionKey, hmacKey, and signingKey explicitly and store them in the platform keystore.

See MnemoPayConfig in src/types/index.ts for optional overrides.

Memory embeddings

MemoryStore / EncryptedSync use one async embedder, configured on MnemoPayConfig:

Install semantic backend when you need it:

npm install @xenova/transformers

MnemoPay.create({
  agentId: 'agent-1',
  embeddings: 'semantic',
  embeddingDimensions: 384,
});

LongMem eval (memory scale + recall)

npm run eval:longmem              # default hash embeddings
npm run eval:longmem:semantic     # same benchmark with Xenova MiniLM (peer dep installed)

Examples:

LONGMEM_N=1000 npm run eval:longmem
LONGMEM_N=5000 LONGMEM_SAMPLES=64 LONGMEM_RECALL_LIMIT=60 npm run eval:longmem

The benchmark resets the in-process memory write rate limiter every 200 retains so LONGMEM_N=5000 can finish in one run. Production apps still enforce normal limits.

The eval prints two JSON blocks:

1. exact query — recall text identical to the stored line. With hash embeds this stays near 100% hit@3 at large N unless k is too small; with semantic embeds it should also stay very high for identical strings.
2. paraphrase query — natural-language question referencing the fact index without copying the stored string. Hash embeds yield near-zero hit@5/hit@15; semantic embeds should improve this materially (run npm run eval:longmem:semantic to measure).

Observed locally (hash, default LONGMEM_RECALL_LIMIT): exact hit@3 = 1.0 for LONGMEM_N through 5000; paraphrase hit@5 ≈ 0 (occasional hit@15). Raise LONGMEM_RECALL_LIMIT if exact recall starts missing at huge N.

The first semantic run downloads model weights into the Hugging Face cache (can take a minute on CI — default CI keeps hash-only eval).

This repo’s Jest config uses jest-environment-node-single-context so onnxruntime-node’s instanceof Float32Array checks succeed under Jest (the default VM-isolated environment breaks typed-array identity).

CI

GitHub Actions runs npm test and npm run eval:longmem (with a small LONGMEM_N) on push and pull requests. See .github/workflows/ci.yml.

License

MIT — see package.json.