Template

Research Synthesizer

Subscribes to RSS / arxiv / blogs, filters by your evolving interests, sends a weekly digest with WHY each item matters given your prior reading.

MemoryMulti-channelTools

For: Academics, analysts, founders tracking competitive landscape, and anyone with information FOMO.

Use this templateSee it in action

Research Synthesizer

You have 47 newsletter subscriptions and read zero of them. Hermes reads them all, remembers what you cared about last week, and tells you what changed.

What it does

Hermes is your personal research analyst. It polls RSS feeds, arxiv papers, blogs, newsletters — whatever sources you configure — and filters by your interest model. The interest model isn't static: it evolves based on what you flag as useful, what you ignore, and what threads you ask follow-ups on.

Every week (Sunday afternoon by default), Hermes sends a digest email:

Weekly research digest — Mar 12

Top 5 items this week (out of 412 scanned):

1. Anthropic's new MCP server spec [arxiv] Why it matters to you: you've been tracking the MCP ecosystem since Jan; this is the spec change that affects how Scout's plugin layer interops. Worth 10 minutes.

2. New paper on RAG hybrid retrieval [arxiv] Why it matters: your March 4 conversation thread asked about BM25-vs-vector tradeoffs. This paper benchmarks the hybrid approach you were considering.

... (3 more)

Less critical (38 items batched, full list in attached digest):

  • 12 product-launch announcements (mostly Series A startups)
  • 8 LLM benchmark updates
  • 18 ecosystem chatter items (Twitter discussions of above)

Threads I'm watching for you:

  • Anthropic's MCP roadmap (3 items added this week; thesis still tracking)
  • Local LLM efficiency (1 item added; no thesis change)

When you reply via Telegram, Hermes has the full context. Ask "what was the benchmark number in that RAG paper?" and it answers, because it read it for you.

Why this agent shape

  • Memory — your interest model refines over time. Items you flag as useful raise the weight on similar items in future digests.
  • Multi-channel — digest goes to email (long-form), follow-ups come via Telegram (conversational). Same agent, two channels with the right texture for each surface.
  • Tools — web fetch (read full articles, not just RSS summaries), source-quality scoring (flags ad-driven content vs deep analysis).

What you'll configure

After picking this template, the architect will ask you to:

  1. List your sources. RSS URLs, newsletter forwarding email addresses, arxiv categories. Hermes can suggest starting sources based on your stated interests.
  2. State your active threads. What are you researching right now? (e.g., "LLM observability platforms", "vector DB tradeoffs", "the Anthropic MCP ecosystem"). These bias the filter.
  3. Pick your digest cadence. Weekly default; can do daily for high-volatility topics or monthly for slow-moving research.
  4. Set source quality thresholds. Hermes scores sources over time; you can require min quality = 7/10 to suppress low-signal noise.

What it looks like running

[Telegram, mid-week]
You: "Anything on observability platforms this week?"

Hermes: Three items in the bucket since last digest:

1. Honeycomb published their Q4 'top open issues' retrospective —
   they're explicitly de-emphasizing LLM observability features.
   Confidence: high (their own blog post).

2. Helicone (your competitor watch) launched cost-attribution by
   prompt template. Confidence: high (product launch).

3. A long-form post on observability stack choice from Datadog's
   CTO — strong on traditional metrics, weak on the LLM angle.
   Confidence: medium-high (Datadog is biased toward their own stack).

The Honeycomb retreat is the highest-information item — you've been
tracking whether they enter your market. They're moving the other
direction.

Want me to draft a thread for your competitive intel doc?

When this template isn't right

  • You want push notifications for every item, not batched digests — the architect can swap to per-item alerts, but the value here is the filtering + synthesis, not the raw stream.
  • You're tracking >100 sources — works but the digest gets long. Consider running multiple Hermes instances scoped to topic areas.
  • Your sources are paywalled — Hermes does basic web fetch; needs additional auth setup for paywalled content (talk to the architect).

What you can ask the architect to change

  • "Daily digest at 7am instead of weekly Sunday" — adjust heartbeat
  • "Skip newsletter ingestion, just RSS + arxiv" — narrow the sources
  • "Suggest research questions I haven't thought to ask" — adds an outbound prompt mode where Hermes surfaces meta-questions
  • "Index past digests semantically — let me search across months" — extends the memory with a vector index

See it in action

Sample interactions illustrating how Hermes actually behaves.

Sunday afternoon — the weekly digest fires

You

(no input — heartbeat triggered)

Agent

[Email] Weekly research digest — Mar 12.

Top 5 (out of 412 scanned):

1. Anthropic MCP server spec change [arxiv] — affects Scout plugin layer interop. Worth 10min.
2. RAG hybrid retrieval benchmark paper — directly relevant to your Mar 4 BM25 question. New data on hybrid scoring.
3. Helicone competitor launched cost-attribution by prompt template.
4. Honeycomb de-emphasizing LLM obs features in Q4 retrospective.
5. New paper on KV cache optimization for long contexts.

Threads I'm watching: Anthropic MCP roadmap (3 items added; thesis tracking), local LLM efficiency (1 item; no thesis change).

Less critical: 38 items batched. Full list at the attached digest.

User asks a mid-week follow-up via Telegram

You

What was the benchmark number in that RAG paper from this week?

Agent

From the BAIR paper (Mar 8, https://arxiv.org/abs/2403.xxxxx):

Hybrid retrieval (BM25 + dense embeddings + reciprocal rank fusion) outperformed pure dense on MS MARCO by 4.2 nDCG@10 points and on BEIR by 3.7 nDCG@10. The improvement was largest on out-of-domain queries (+6.1 nDCG) — supports your thesis that hybrid wins when the query distribution shifts.

Key caveat: their RRF weights were tuned per-dataset. The paper doesn't address how to set them without per-dataset data. That's the open question for your work.

User refines interest model

You

Stop sending me anything about open-source LLM benchmarks — I don't care

Agent

Logged. Adjusted interest model: "open-source LLM benchmarks" weight set to -5 (was 0). This will suppress items in that category from future digests.

Specifically: items matching keywords "benchmark", "eval suite", "leaderboard", "MMLU", "HellaSwag" + open-source model names — all filtered out unless they intersect another active thread.

Reverse this anytime by saying "care about benchmarks again."

Ready to deploy this agent?

Picking this template loads the configuration above into the Forge Architect. You can refine anything via plain English before deploying.

Use this template