Building sonn.dev — persistent cross-session memory + a proactive reasoning layer for AI coding agents. Local-first.

devto

Aleksandar Kovacevic  • Peerlist

Sonn is a local-first memory and reasoning layer for AI coding agents, built primarily for Claude Code users. Most tools in this space store text and hand it back when asked — that's retrieval. Sonn goes further: it steps in before the mistake, steering the agent proactively instead of just remembering what happened.The problem. AI coding agents are brilliant inside a session and amnesiac the moment you close it. Developers re-explain the same context, re-fix the same bugs, and hand-maintain <a target="_blank" rel="noopener noreferrer nofollow ugc" class="break-all link-break text-gray-500 hover:underline" href="http://CLAUDE.md">CLAUDE.md</a> files that go stale by the third session. Even Anthropic closed the community request for persistent memory — the "goldfish" issue — as not planned.How Sonn works — three layers over one engine:<ul><li>Whispers — trigger-gated behavioral nudges injected into the session. Instead of only adding to what the model knows, whispers change how it behaves — the closest thing to fine-tuning without access to the weights.</li><li>Angel — a persistent reasoning process that observes, forms opinions about what matters, recognizes recurring patterns, validates them across independent agents, and injects the right guidance before you ask.</li><li>Memory — the local-first foundation. Your code and memory stay on your machine; only the reasoning layer runs server-side.</li></ul>Competitors make the agent know more. Sonn makes the agent get better. That's the difference between a filing cabinet and a mentor.Local-first by architecture. Your code never leaves your machine - generation runs through your own Claude credentials, inside your own trust boundary. Only the reasoning core is hosted, which solves three problems at once: privacy - nothing to route through vendor cloud, IP protection - the reasoning engine never ships to the client, so there's nothing to extract. Status. The engine is proven, not a slide. It runs daily across several genuinely different real projects — including building Sonn itself — and scores 90.6% on the LongMemEval memory benchmark with a local model. Team. Built by a two-enthusiast team from Serbia, we hope to scale team soon:)