Telepath: A Memory-Aware Visualizer That Already Knows You
You open a visualization tool and type "show me how I've been spending my time." The tool asks: what time range? What breakdown? What data source? By question four, you close the tab.
Telepath (4 stars, MIT) was built to skip those questions. It reads your Hermes Agent's persistent memory first — ~/.hermes/memories/USER.md and MEMORY.md — and fills in the blanks from what it already knows about you. The more Hermes has learned, the quieter Telepath gets.
Built for the Hermes Agent Creative Hackathon (Nous Research × Kimi/Moonshot, May 2026), Telepath is a weekend project that punches above its weight. It is a Next.js 16 app that takes vague intent and produces Vega-Lite charts, Mermaid diagrams, or single-screen infographic slides. All driven by Kimi K2.
How It Works
Telepath runs three structured prompts against the Kimi K2 API:
parse_intent — takes your vague request, restates the goal in concrete terms, picks the output format (chart / diagram / slide), and grounds every renderer dimension. For each dimension — audience, time window, data source, breakdown method, color palette — it tags the source: memory (found in your profile), default (reasonable guess), asked (needs a question), or missing (no clue).
pick_question — ranks unresolved dimensions by impact times confidence. Asks at most one question, presented as clickable chips. If everything was resolved from memory, it skips entirely.
synthesize_spec — emits a Vega-Lite v5 JSON spec, a Mermaid diagram source, or a slide layout JSON. The renderer picks up the spec and draws.
After a successful render, the resolved intent-to-spec pair gets written back to ~/.hermes/skills/telepath/<slug>/DESCRIPTION.md. Next time you ask the same thing, Telepath replays from the saved spec. Zero LLM round-trips.
The Demo That Shows It Working
The project ships a demo fixture with two modes:
Cold start — Telepath has no memory to read. Ask "chart how I've been spending my time" and it asks 1–2 chip questions. Renders.
Memory active — toggle off cold start. Same question. Telepath pulls your time window and breakdown preference from USER.md, glows the chips it used in the left rail, asks zero questions. Renders.
Replay — click a saved skill in the right rail. Instant re-render, no API calls.
It is a tight demo. The UX is cleaner than most hackathon projects: a knowledge rail on the left showing which memory fields were used, a skills rail on the right showing saved visualizations, the composer at center. Three-column layout that feels like a real product.
The Architecture
The stack is modern: Next.js 16 App Router, React 19, Tailwind 4, pnpm. The Kimi K2 integration uses the Moonshot OpenAI-compatible API. Vega-Lite handles charts. Mermaid handles diagrams. Custom slide primitives handle infographics.
No database. The Hermes filesystem is the database. Memory reads come from ~/.hermes/memories/. Skill writes go to ~/.hermes/skills/telepath/. Session data can optionally come from ~/.hermes/sessions/ and cron metadata from ~/.hermes/cron/.
The lib directory tells the real story. There are modules for Hermes memory parsing, session extraction, cron metadata, Kimi API calls, and a skill generalization pipeline that can distill a resolved visualization into a reusable skill. It is more than a demo — it has the skeleton of a persistent visualization knowledge base.
What Makes It Interesting
Telepath is not a general-purpose charting tool. It is a charting tool that reads your agent's brain. That distinction matters.
Hermes already accumulates durable information about you: preferences, projects, recurring tasks, communication patterns. That information sits in markdown files in ~/.hermes/memories/. Most visualization tools ignore it. Telepath treats it as ground truth.
The implication: as your agent gets smarter about you, your visualization tool asks fewer questions. The two systems compound. That is a design pattern worth watching.
The hackathon context also matters. Nous Research partnered with Kimi/Moonshot for this event, and Telepath qualifies for both the Main and Kimi tracks. If it places, expect more polish and possibly official Hermes integration.
The Limitations
Four stars. Four days old (initial commit: May 2, 2026). Hackathon code.
The memory grounding is only as good as your USER.md and MEMORY.md. If those files are sparse — and for most Hermes users they are — Telepath falls back to asking questions, which defeats its core pitch. The "it already knew" tagline depends on Hermes having done its job first.
Kimi K2 is the only supported model. The Moonshot API is the only provider path. If you do not have a Moonshot API key, Telepath is a static demo.
The output surface is limited to charts, diagrams, and single slides. No dashboards with multiple panels. No interactive drill-down. No data export beyond the saved skill. For a hackathon project that is fine. For a daily driver, it would get cramped fast.
There is no Hermes skill for invoking Telepath from inside the agent. You run it as a separate dev server and paste results back. The integration is filesystem-based — Telepath reads Hermes's output directory — but there is no bidirectional tool interface.
Should You Install It?
Not yet. Four stars and four days of development do not make a production tool.
But watch it. The idea — reading agent memory to reduce prompt friction in downstream tools — generalizes beyond visualization. Any tool that needs user context (search refinement, notification filtering, meeting prep) could benefit from the same memory-grounding pattern. Telepath is an early implementation of a design pattern that will matter.
What would make it installable: a proper Hermes skill with tool invocation, multi-model support beyond Kimi, and a dashboard view for composing multi-panel visualizations. Until then, it is a well-executed experiment.