We just hit 81.3% on LongMemEval, the standard benchmark for agent memory. Now we're aiming to bring this to @SolanaMobile Seeker. Here's what we built, what we learned, and why mobile agents need a memory layer more than desktop ones do.

The default approach to agent memory is "just use vectors." Embed everything. Search by cosine similarity. Return top-k. Doesn’t cut it anymore in 2026 Ask "what backend are we using?" and vector search returns every conversation about technology. It doesn't find "we switched to Supabase last Tuesday." Pure semantic search can't find specific facts buried in thousands of memories.

The fix wasn't a bigger model or more tokens. It was a better retrieval pipeline. Five changes: 1. BM25 keyword search alongside vectors - catches exact names, numbers, addresses that embeddings miss 2. Reciprocal Rank Fusion to merge both result sets 3. Cross-encoder reranking - rescores each candidate against the actual query 4. Temporal bonds - time-aware linking so "what happened last week" works 5. Confidence gating - uncertain? say so. don't hallucinate None of these are novel. All well-studied in IR. The hard part is making them work together at low latency for real-time agents.

Why Seeker specifically? Seeker already solved the two hardest problems for agent memory: Identity: Seed Vault gives you a hardware-secured wallet on the device. That wallet becomes your memory identity. No accounts. No OAuth. No passwords. Payments: SKR makes $0.001/query micro-transactions viable. Every memory store, recall, and verification is a micro payment. The economics only work on a chain where sub cent transactions aren't a joke 100K+ Seeker users already have wallets. Every AI app on the dApp Store could share one memory layer; scoped to your wallet, portable across apps. Your DeFi copilot knows your risk tolerance because your AI assistant already learned it. One wallet. Every app remembers you.

What's live today: Cortex API, npm + Python SDK (5,000+ installs), MCP integration, Solana on-chain verification, 1.3M+ memories stored, 600+ active agents. What we're going to be building next with @SolanaMobile: Mobile Wallet Adapter integration, Seed Vault identity, SKR payment rails, offline-first cache, and a reference app on the dApp Store. The memory engine works. The Seeker-native layer is what's next. We applied for the Solana Mobile Builder Grant to make it happen. We publish real production scores because this space needs more honesty and fewer marketing benchmarks.

Shared memory infrastructure for every app on Seeker. One wallet, persistent context, zero cold starts. Open source. $0.001/query. Building in public. @solanamobile @toly @solana_devs