D_AI — memory and decisions for long projects

On long projects run with AI agents, every session starts cold: why we chose this, what we tried, what's left. When an agent doesn't know, it guesses and carries on as if it did.

I didn't build another model. I built a state and policy layer over the host's own model loop. Work starts from a charter with an objective, a deliverable, an acceptance criterion and a budget, and gaps get reported instead of filled with assumptions. A task moves through 17 states and only reaches COMPLETED through a recorded attempt, and decisions are never edited: a new one supersedes the old and names who authorised it.

Memory lives in three separate stores, person, project and global, and every record has its own encryption key, so a permanent purge destroys the key itself. An OpenCode adapter ships agents, skills, commands and an MCP server, next to specs for ten specialist systems from research to strategy. Model routing has one blunt rule: unknown cost is not zero.

The repo is public and ships an honest implementation status: the core is covered by tests (193 passing at the repo's own checkpoint) and per-domain professional validation hasn't happened yet. A first A/B test, same model and same task, scored the system arm higher in two rounds. The write-up itself says two rounds are not an experiment, and that the core's tools were never called in either. That's the next fix.