Skip to main content

Hermes Agent on Windows: Set Up Persistent Memory with Hindsight

· 7 min read
Ben Bartholomew
Ben Bartholomew
Hindsight Team

Hermes Agent on Windows with Hindsight

Earlier this week, Nous Research announced that Hermes Agent is now natively supported on Windows. Here's how to give it persistent memory.

Hermes Agent is a self-improving assistant with 40+ tools — chat, research, voice, gateway integrations into Telegram, Discord, Slack, and more. Out of the box, every conversation starts from zero: Hermes doesn't remember what you discussed yesterday, what project you were planning, or what preferences you've already shared. Hindsight closes that gap. Once configured, Hermes recalls relevant facts before every reply and retains the conversation afterward — across sessions, across platforms, across restarts.

This post is the platform-neutral Windows guide. If you specifically want a coding assistant with codebase memory, see the Hermes Coding Assistant on Windows companion.

What You Get on Windows​

The full Hermes + Hindsight feature set, native:

  • Three deployment modes — pick Cloud, Local Embedded, or Local External in the setup wizard. No Docker, no WSL, no separate PostgreSQL install. Local Embedded ships with pg0 (an embedded Postgres distribution); first boot unpacks the binaries and runs initdb, subsequent starts are fast.
  • Auto-recall and auto-retain — every conversation starts with relevant memories injected into Hermes's context, and every conversation retains what you discussed for the next one.
  • Cross-platform identity — if you run Hermes Gateway (Telegram, Discord, Slack), all platforms share the same memory bank. Tell Hermes something in Telegram, ask about it later from Discord, and it remembers.

The only Windows-specific gotcha is character encoding (more on that below).

One-Command Setup​

There's no Windows-specific install dance. From PowerShell:

hermes memory setup

Pick hindsight from the provider list. The wizard then asks you to choose a mode:

Select mode
↑↓ navigate  ENTER/SPACE select  ESC cancel

→ (●) Cloud           Hindsight Cloud API (lightweight, just needs an API key)
  (○) Local Embedded  Run Hindsight locally (downloads ~200MB, needs LLM key)
  (○) Local External  Connect to an existing Hindsight instance

What each mode asks for:

  • Cloud — your hsk_... token from ui.hindsight.vectorize.io/connect. The default API URL (https://api.hindsight.vectorize.io) is correct — just accept it.
  • Local Embedded — the wizard downloads and runs Hindsight locally (~200MB on first install, including the embedded PostgreSQL binary). You'll be prompted for an LLM provider (OpenAI, Anthropic, Gemini, Groq, Ollama, etc.), an API key for that provider, and an optional model override. Nothing else to install — uvx handles fetching hindsight-embed, and the daemon spawns automatically.
  • Local External — point Hermes at a Hindsight instance you're already running (your own server, a teammate's, a self-hosted box). Provide the API URL and, if the instance requires it, an API key.

Confirm it's wired up:

hermes memory status

You should see provider: hindsight and status: ready. On Local Embedded, first boot takes 60–90 seconds while pg0 unpacks Postgres and runs initdb.

What Hermes Remembers​

After a few conversations, the kind of facts that enter memory automatically:

  • "User prefers metric units and 24-hour time."
  • "Planning a camping trip to Wyoming in August — group of six, mix of tent and cabin."
  • "Weekly standup notes go to #team-eng on Slack at 5 PM ET."
  • "User's writing voice is concise, avoids exclamation points, and prefers em-dashes."

You don't tell Hermes any of this directly. The write pipeline extracts it from your conversations — the questions you ask, the decisions you make, the context you share along the way.

Three Workflows Where Memory Pays Off​

Picking Up Where You Left Off​

Without memory, every session starts with re-explaining. You spent yesterday planning a trip; today you open Hermes and say "let's keep going" — and it has no idea what you mean.

With Hindsight, the next session opens on the actual continuation:

"Where did we land on the day-three itinerary for the Wyoming trip?"

Hermes already knows the trip, the group size, the constraints you've discussed, and the open questions. The first message is real work, not setup.

Talking to Hermes Across Platforms​

Hermes Gateway runs across Telegram, Discord, Slack, and other messaging platforms. Without memory, each platform feels like a different assistant — you tell Hermes a fact in Telegram on Monday, ask about it from Discord on Wednesday, and the Discord-side Hermes shrugs.

With Hindsight, the memory bank is shared. The plugin hooks every turn on every platform — whatever you say in one channel becomes recallable from any other. One assistant, many surfaces.

Long-Running Research and Writing​

Multi-week projects — a research dive, a book outline, a job hunt — accumulate hundreds of facts. Without memory, you fight the agent for context every session: re-pasting the brief, re-explaining who the audience is, re-stating constraints you set three weeks ago.

With memory, you query what Hermes already knows:

What did we decide about the audience for the agent-memory ebook?
What's still open on the job hunt — which roles am I waiting to hear back on?

The answers come from prior sessions — including the small decisions and side-notes that never made it into a doc.

Which Mode Should You Pick?​

A quick decision tree:

  • Solo, want zero ops: Cloud. One API key, no daemon to think about, memory accessible from any machine you log into.
  • Solo, privacy-sensitive or offline-heavy: Local Embedded. Memory bank stays on your machine, works on flights and air-gapped networks (you still need the network for the LLM call itself).
  • Team with a shared Hindsight server: Local External. Point each user's Hermes at the same Hindsight instance and share a bank ID — institutional knowledge accumulates across the team.

The Hindsight v0.5.5 release added native Windows support for embedded mode, and every subsequent release has been verified on Windows. The v0.4.21 native Windows post has the original walkthrough for running without Docker.

Common Windows Gotchas​

A few things to watch for, distilled from the Windows CI runs and user reports:

  • UTF-8 encoding — Hermes and Hindsight log status with checkmarks and box-drawing characters (✓, ─, │) that crash Windows's default cp1252 codec. If you see a UnicodeEncodeError on first run, set PYTHONUTF8=1 and PYTHONIOENCODING=utf-8 in your PowerShell profile. The Hindsight Windows smoke test sets both on every run for the same reason.
  • First-boot Postgres init time (Local Embedded only) — 60–90 seconds on a cold Windows machine while pg0 unpacks Postgres and runs initdb. Logs land at ~/.hindsight/profiles/<profile>.log.
  • Long file paths — if your home directory or workspace lives deep in the tree, enable LongPathsEnabled in the registry. Hindsight itself stays under the 260-char limit, but some Python wheel installs don't.
  • Antivirus on the daemon — Windows Defender occasionally flags the embedded Postgres binary on first unpack. Add ~/.hindsight/ to the exclusion list if you hit slow startup or quarantine warnings.

The Longer You Use It, the Less You Explain​

Session one, Hermes knows nothing about you. Session five, it knows the units you prefer, the projects you're juggling, the people in your life. Session 30, it knows the recurring tasks, the writing voice, the small preferences that never get documented but always trip up a fresh assistant.

Hermes just shipped native Windows. Hindsight has been there. Plug them together and your assistant gets a memory on the platform you already work on.

Set it up with hermes memory setup, or start with the Hindsight Windows installation guide.

Further reading: