Save Slack and Whatsapp Conversations into Hermes + Gbrain Longterm Memory

You converse with your Hermes Agent via Whatsapp & Slack. Here’s how you can save them into longterm memory in Hermes + GBrain.

Resources:

This is an article in the series Hermes + Gbrain. You can find earlier posts here:

Hermes is session-based. Every conversation starts fresh.

GBrain gives it long-term memory, but only if you explicitly feed it. By default, your Slack and WhatsApp conversations with Hermes are stored in a local SQLite database (state.db) and never make it into GBrain. 130 sessions sitting there. Zero intelligence extracted.

I fixed this with a nightly distillation pipeline. Every night at 11pm, the day’s conversations are read from state.db, distilled by GPT-4o into decisions / action items / observations / open questions, and written to GBrain as structured knowledge pages. By morning, Hermes has full context on everything discussed the day before.

Here’s exactly how I built it.

The Problem

Hermes stores all conversations in ~/.hermes/state.db. Check yours:

sqlite3 ~/.hermes/state.db "SELECT source, COUNT(*) FROM sessions GROUP BY source;"

You’ll see something like:

slack|75
whatsapp|56
cli|59
cron|388

Those 131 Slack + WhatsApp sessions contain real decisions, action items, trip plans, business ideas — and none of it is in GBrain. Hermes forgets all of it the moment the session ends.

Gotcha: GBrain’s memory.provider in ~/.hermes/config.yaml defaults to blank. This means cross-session memory only exists if you explicitly build it. The session store is not the same as long-term memory.

The Architecture

state.db (Slack + WhatsApp sessions)
↓  collect_conversations.py  [deterministic — no LLM]
digest-YYYY-MM-DD.md
↓  distill_to_gbrain.py  [GPT-4o — judgment layer]
daily-YYYY-MM-DD.md + entity pages
↓  gbrain put <slug>
GBrain long-term memory

Two scripts. One shell wrapper. One cron job.

The pattern is identical to GBrain’s existing recipes (email-to-brain, x-to-brain): deterministic code handles data extraction, LLM handles judgment. Never mix them.

Prerequisites

Hermes Agent running on AWS VPS (Ubuntu)
GBrain installed at /home/ubuntu/.bun/bin/gbrain
OpenAI API key configured in GBrain (gbrain config get openai_api_key)
Python at /home/ubuntu/.hermes/hermes-agent/venv/bin/python
sqlite3 installed (sudo apt install sqlite3 -y)

Part 1: The Collector

collect_conversations.py reads today’s sessions from state.db and writes a clean markdown digest. No LLM involved — pure SQL + text formatting.

Install at:

~/.hermes/scripts/collect_conversations.py

Key design decisions:

Filters to source IN (‘slack’, ‘whatsapp’) only — skips cron noise
Strips the Hermes persona boilerplate from system_prompt (it’s 2,000 characters of SOUL.md on every session — not useful)
Skips sessions with no real user messages
Truncates individual messages at 2,000 characters to handle file uploads
Splits JSON content blocks (tool calls) and extracts only readable text

Test it:

/home/ubuntu/.hermes/hermes-agent/venv/bin/python \
      ~/.hermes/scripts/collect_conversations.py \
       --date 2026-05-10 \
       --output /tmp/digest-test.md
cat /tmp/digest-test.md | head -60

Expected output

# Hermes Conversation Digest — 2026-05-10
Total sessions: 7
---
## Session 1: WhatsApp Persona Config
**Source:** SLACK  |  **Started:** 18:26
**[18:26] You:** Hermes ⇔ WhatsApp Persona Config...
**[18:26] Hermes:** Got it. Send the next bit of context when ready.

Part 2: The Distiller

distill_to_gbrain.py reads the digest and extracts structured knowledge using GPT-4o. It writes one daily page + individual entity pages to GBrain.

Install at:

~/.hermes/scripts/distill_to_gbrain.py

Model matters. I initially used gpt-4o-mini. The extractions were thin — 2 generic action items from a day with 25 sessions. Switched to gpt-4o. Night and day difference.

Chunking matters. A busy day can produce 37,000+ character digests. GPT-4o supports 128k context, but I chunk at 40,000 characters at session boundaries to keep each call focused. Results from multiple chunks are merged and deduplicated.

The distillation prompt extracts:

Decisions — things resolved or committed to
Action items — concrete next steps with owner
Observations — insights, learnings, what worked/didn’t
Open questions — raised but unresolved
Entities — real people and companies with context

Gotcha: LLMs left unconstrained will create entity pages for email addresses, stock ticker symbols, tool names, and groups of people concatenated together. Add strict rules to the prompt:

NEVER create entity entries for:

- Email addresses

- Stock ticker symbols (AAPL, NVDA)

- Tools or platforms mentioned in passing (WhatsApp, Slack, pdfplumber)

- Groups of people combined into one name

- Sudhir himself — he is the user, not an entity

- Vague single names with no surname or role

Gotcha: gbrain import <file> expects a directory, not a file path. Use gbrain put <slug> with stdin instead:

result = subprocess.run([gbrain, “put”, slug], input=page_content, capture_output=True, text=True, env=env)

Test with dry run:

export OPENAI_API_KEY=$(/home/ubuntu/.bun/bin/gbrain config get openai_api_key)
/home/ubuntu/.hermes/hermes-agent/venv/bin/python \
~/.hermes/scripts/distill_to_gbrain.py \
--digest /tmp/digest-test.md \
--date 2026-05-10 \
--dry-run

cat ~/.hermes/distill/pages/raw-2026-05-10.json

Expected output:

{
   "date": "2026-05-10",
   "decisions":
   [
      {
         "title": "WhatsApp agent configuration confirmed",
         "detail": "require_mention: true confirmed to minimise unsolicited responses."
      }
   ],
   "action_items": 
   [
      {
        "task": "Investigate false-positive safety block in daily morning brief job",
        "owner": "Hermes",
        "context": "Job was blocked, not a WhatsApp delivery failure."
      }
   ],
    ...
}

Part 3: The Nightly Pipeline

nightly_distill.sh orchestrates the full flow: collect → distill → gbrain put.

Install it using:

~/.hermes/scripts/nightly_distill.sh

It automatically pulls the OpenAI key from GBrain config — no manual key management:

OPENAI_API_KEY=$("$GBRAIN_BIN" config get openai_api_key 2>/dev/null)
export OPENAI_API_KEY

Part 4: Backfill History

Before setting up the nightly cron, backfill everything. backfill_distill.sh loops over a date range, processes each day, and runs gbrain embed –stale once at the end.

bash ~/.hermes/scripts/backfill_distill.sh --days 21
── 2026-05-09 ──────────────────────────────
✅ Collected 9 sessions
✅ Distilled → GBrain
── 2026-05-10 ──────────────────────────────
✅ Collected 7 sessions
✅ Distilled → GBrain
════════════════════════════════════════
Backfill complete
Days processed : 22
Days skipped   : 1 (no conversations)
Total sessions : 130
Failed         : 0
════════════════════════════════════════
▶ Running gbrain embed --stale ...
Embedded 18 chunks across 14 pages
✅ Embeddings updated

Gotcha: If the backfill gets interrupted (Ctrl+C), just re-run. Each day is idempotent — gbrain put overwrites existing pages cleanly.

Part 5: Install the Nightly Cron

crontab -e

0 23 * * * /home/ubuntu/.hermes/scripts/nightly_distill.sh >> /home/ubuntu/.hermes/distill/logs/nightly.log 2>&1

Part 6: The Remember Skill

Check it ran the next morning:

cat ~/.hermes/distill/logs/nightly.log

For things too important to wait until 11pm, add an instant-save skill. Send Remember: from Slack or WhatsApp:

Remember: We decided to use HIFO cost basis for the CA tool
Remember: Sam's number is +1-732xxx — call him about the lease

remember.py
writes the note directly to GBrain and embeds it immediately — searchable within seconds.

~/.hermes/skills/productivity/remember/scripts/remember.py

Verify It Works

/home/ubuntu/.bun/bin/gbrain query "Toronto trip"
/home/ubuntu/.bun/bin/gbrain query "AI automation"
/home/ubuntu/.bun/bin/gbrain query "WhatsApp morning brief"

Top result for each should be the relevant daily page with score > 0.75.

Cost

Component                                     Monthly
CostGPT-4o distillation (~10k tokens/day)    ~$0.75/mo
OpenAI embeddings (text-embedding-3-large)   ~$0.10/mo
Total                                        < $1/mo

What’s Now in GBrain

Every night:

All Slack + WhatsApp conversations → structured knowledge
Decisions, action items, observations indexed and searchable
Entity pages for real people and companies, updated with timeline entries
Remember: available for instant saves anytime

Hermes wakes up each morning knowing what you decided, planned, and discussed the day before. The brain compounds.

Resources:

This is an article in the series Hermes + Gbrain. You can find earlier posts here:

Thanks to @Teknium and the @NousResearch team for building Hermes, and @garrytan for open-sourcing GBrain.

That’s all folks!

If you have suggestions to improve this further, please comment. Would love to learn. Please Like, Repost so other Hermes + GBrain users can benefit.

Please Follow to learn with me how to Build the World’s Best Personal AI Assistant for yourself.

Category: Artificial Intelligence

Save Slack and Whatsapp Conversations into Hermes + Gbrain Longterm Memory

Resources:

The Problem

The Architecture

Prerequisites

Part 1: The Collector

Part 2: The Distiller

Part 3: The Nightly Pipeline

Part 4: Backfill History

Part 5: Install the Nightly Cron

Part 6: The Remember Skill

Verify It Works

Cost

What’s Now in GBrain

Resources:

Like this:

Related

Leave a ReplyCancel reply

Resources:

The Problem

The Architecture

Prerequisites

Part 1: The Collector

Part 2: The Distiller

Part 3: The Nightly Pipeline

Part 4: Backfill History

Part 5: Install the Nightly Cron

Part 6: The Remember Skill

Verify It Works

Cost

What’s Now in GBrain

Resources:

Share this:

Like this:

Related

Leave a ReplyCancel reply