Engram, a psychology and neuroscience based learning plugin, is now on DeepSeek Harness (it teaches you, not the agent)

Agentic AI made building about 10x faster. Learning didn't get any faster. I noticed I was shipping systems I couldn't re-explain a week later, and it started to bug me. We have a 10x tool for building, so I wanted the equivalent for learning, in the same terminal where the building happens. That's Engram. It's a tutoring loop grounded in the memory research, and it's aimed at you, not the agent. dsh's whole premise is that everything is a plugin: models, tools, skills, the UI. This is a plugin for the one part that doesn't ship, the person running it. - It breaks a topic into a first-principles concept graph and teaches one node at a time. - It won't explain anything until you've committed to a guess first. Retrieval before instruction is the single best-replicated result in learning science, and also the part every chatbot skips because agreeing with you is easier. - Your recall gets graded by a separate blind assessor that never sees the tutoring conversation, so the tutor can't inflate grades on its own teaching. - Reviews are scheduled with FSRS, so they show up right before you'd forget. A few minutes a day. - Wrong models get logged verbatim and re-probed later. Mine has ten entries for transformers alone. Honest origin story: with an early version I encoded seven concepts, never came back, and lost half of them right on schedule. The whole loop got redesigned around returning. No streaks, no XP. v1.13 makes dsh the eighth platform (after Claude Code, Codex, OpenCode, Hermes, Antigravity, OpenClaw, pi), and it's the thinnest port yet: zero adapter code, because dsh natively reads SKILL.md folders, AGENTS.md, and Claude Code hooks. Everything was verified against a real dsh 0.1.0-rc.6, except a full tutoring session, since dsh has no free models and I stopped at the key wall. If you run one and something's off, open an issue. Install (needs a DEEPSEEK_API_KEY): git clone github.com/nagisanzenin/engram ~/.agents/engram mkdir -p ~/.agents/skills for s in learn review coach; do ln -sfn ~/.agents/engram/skills/$s ~/.agents/skills/$s; done Then type /learn in dsh. The optional session-start nudge and a guarded version of that loop are in the walkthrough: github.com/nagisanzenin/engram/blob/main/INSTALL-DSH.md Repo: github.com/nagisanzenin/engram — your learning state is local JSON shared across all eight tools, so a topic encoded in Claude Code gets reviewed here. I'm the author, so grain of salt. But I've been dogfooding it daily to learn transformer internals and it's the first setup where week-old material actually stays with me.