Local-first AI,
that actually remembers you

Every AI framework vendor claims their agents are reliable. There's no third-party number to challenge or confirm that. The companies best positioned to build this benchmark (Patronus, Braintrust, LangSmith) sell into the same framework market, so publishing "this framework collapses to wrong answers" antagonizes their customers. The space needs a measurement entity that doesn't sell into it.

Built a deterministic scoring pipeline (no LLM in the grading loop) across multi-agent convergence and AI memory recall. Every result is an Ed25519-signed JSON receipt with methodology hash, fixture SHA, raw scores, and environment pin, committed to a public ledger. Anyone re-runs, anyone verifies. The vendor-cert tier turns the same harness into a third-party performance attestation framework vendors can publish on their own site.

Most chat apps lock you into one model and forget everything between sessions. Personal assistants don't know your business; enterprise AI tools don't respect your privacy. Teams want both, plus context that actually survives.

Built a cross-platform Electron app with a unified inference engine that speaks llama.cpp (local) and provider APIs (OpenAI, Anthropic, OpenRouter), an IPC bridge to Engram for memory and Persona for identity, and a per-conversation provider switch. Personal tier ships free; Pro adds cross-device continuity for agents, memory, and personality plus advanced multi-agent orchestration; Enterprise ingests company docs, code, calendars, and tickets with on-prem deployment, SSO, and audit logging.

AI assistants forget every conversation the moment the context window fills. Each session starts from zero, with the same questions, same explanations, same setup work. And memory that lives only on one machine can't follow you between laptop, desktop, and phone.

Built a 9-stage hybrid retrieval pipeline backed by LanceDB vectors plus a typed knowledge graph. Pre-compaction handoff hooks dump session state to disk before context windows fill. Procedural rules surface based on emotional valence, recency, and outcome feedback. The hosted Engram Cloud tier syncs the same memory across every device you use, with end-to-end encryption and team-shared memory spaces.

Out-of-the-box LLM agents speak in a corporate-helpful monotone. Any personality you give them has to be re-prompted each conversation, and there's no path for them to grow into how you actually communicate.

Five-layer system covering personality, communication style, working style, a soul library, and behavioral signals. Persona reads user reactions (correction, approval, frustration, praise) and proposes evolutions over time with an explicit accept/reject flow so the agent doesn't drift unsupervised.

Work knowledge sprawls across 8-15 SaaS tools. Glean and Microsoft Copilot for Work solve this only if your data lives in their cloud, which is a non-starter for legal, finance, healthcare, and any company with real data-sovereignty requirements. Permission-aware retrieval across all those sources is the missing piece nobody open-source has nailed.

Built an MCP server with workspace-scoped tenancy, hybrid retrieval, and ACL mirroring enforced at the database layer (Postgres RLS). Strips the bundled-LLM dependency in favor of an MCP enrichment-callback protocol. Your client (Pyre, Claude Desktop, anything MCP-speaking) supplies the LLM. Cortex supplies the memory and the connectors. The enterprise wedge of the OneNomad stack.