Hermes Agent is an agent that does work and keeps its own memory. MemPalace is a memory store that an agent reads from and writes to. A Hermes model is just a model.
Those names get compared because both projects promise that your AI will remember. But a searchable record of last month's work and an assistant that runs tomorrow morning's task solve different problems.
For a small business, start with the job that keeps slipping. Do you need something that acts, something that finds old work, or just a better note? Adding one more system to babysit is easy. Finding the decision you made last Tuesday is the part that pays.
This comparison uses primary sources read on October 29, 2026. We have not installed Hermes Agent or MemPalace for this article. Our existing Hermes model setup is described below. The older MemPalace explainer and Hermes Agent article provide background; this is the current comparison, including corrections to early claims.
First, separate the three things
Scroll to compare columns and rows when the table extends beyond the box.
| Question | Hermes Agent | MemPalace | Hermes model |
|---|---|---|---|
| Kind of tool | Agent runtime that calls tools and manages tasks | Memory storage and retrieval software | Language model weights |
| What it remembers | Curated notes, user preferences, session history and reusable skills | Imported conversations, project material and memories filed by agents | No durable personal memory by itself |
| Where data lives | Local files and session database; optional external providers | Local database by default; optional remote components | Wherever inference runs; the caller supplies context |
| What it needs | Supported computer, installer and a model connection | Python, storage backend, embedding model and an agent to use it | Model runner, enough memory and a suitable chat template |
| Who it fits | Operator who wants repeatable work done through tools or chat | Developer who needs past work searchable across agents | Operator choosing the model behind an application |
| Main risk | Tool permissions turn a bad instruction into an action | Sensitive or misleading material becomes reusable context | Fluent output mistaken for verified knowledge |
An embedding model turns text into numbers that help find similar text. It does not answer your business question. The answering model still has to read what retrieval returns and use it correctly.
Hermes Agent: the worker, with a notebook
Nous Research launched Hermes Agent on February 25, 2026. It is MIT licensed. The latest GitHub release we found was v0.22.5, published October 23 UTC (its notes are dated October 22). The release before it came out two days earlier. Plan on updating often.
You can use its terminal interface or connect messaging gateways such as Telegram, Discord, Slack, WhatsApp and Signal. Its scheduler can run a task once or repeatedly and deliver the result to a chat. It can also delegate work to child agents, each with its own conversation context. Separate conversations do not establish separate security boundaries.
Think of a repair shop owner asking for tomorrow's job list from a phone. Hermes supplies the agent loop, tools, schedule and reply channel. The owner still needs to connect the actual booking system and decide which actions require review.
Installation and model choice
Use the official installation instructions. The current source installer manages Python 3.14 and supporting tools. Its wider package metadata range exists partly to let older installations update; it is not the recommended runtime. Linux, Apple Silicon macOS and native Windows are supported, with platform qualifications in the support matrix. Windows is no longer a WSL only story.
Read the installer before running it, then use hermes setup and hermes model to configure the agent. You do not need a GPU if the language model runs elsewhere.
The provider documentation covers Nous Portal, OpenRouter, Anthropic, OpenAI and other services. Ollama and compatible endpoints can supply local inference. API compatibility means the programs can communicate. It does not prove that your chosen model will make reliable tool calls.
The software license is free. Model calls, paid tools, hosted sandboxes, electricity and your maintenance time still cost something. When you choose a remote model, the prompt and included tool results leave the machine. A messaging gateway adds the messaging provider. Optional speech, search and memory services can add more destinations.
What its memory actually contains
Hermes has several distinct memory mechanisms:
MEMORY.mdholds compact environmental facts and lessons.USER.mdholds preferences and information about you. These are curated notes loaded into the prompt, not the entire transcript.- A local SQLite database keeps sessions searchable. Its full text search can recover earlier messages without pretending that every conversation fits in the current prompt.
- Skills record reusable procedures. The agent can write and improve them after doing work. Only the saved instructions change. The model's weights stay the same.
Optional memory providers include Honcho, Mem0, Hindsight, OpenViking, Holographic, RetainDB, ByteRover, Supermemory and Memori. Packaging and maintenance differ between them. Honcho adds richer user modeling, for example. One external provider can be active alongside the built in memory. Check where that provider stores data before enabling it.
A Hermes model is a separate model choice
The Hermes model family consists of fine tunes: open weight base models trained further by Nous. Downloading those weights does not install the agent, a scheduler or a personal memory database.
The official model cards identify these variants. B means billions of parameters, not gigabytes of required RAM.
Scroll to compare columns and rows when the table extends beyond the box.
| Family | Bases and sizes | License |
|---|---|---|
| Hermes 3 | Llama 3.1 at 8B, 70B and 405B; later Llama 3.2 at 3B | Llama license terms |
| Hermes 4 | Llama at 70B and 405B; Qwen 3 at 14B | Llama terms for Llama variants; Apache 2.0 for 14B |
| Hermes 4.3 | Seed OSS at 36B | Apache 2.0 |
Sources: Hermes 3 report, 3B card, Hermes 4 70B, 405B, 14B and 4.3 36B. Hermes 4.3 is the newest Hermes model we could verify in Nous's release catalog.
The cards describe tool calls as JSON containing a function name and arguments inside <tool_call> tags. Chat templates differ: Hermes 3 uses ChatML; the larger Hermes 4 cards show Llama style role headers. The serving software must parse the call, run an allowed tool and return its result. Model weights alone do none of that bookkeeping.
MemPalace: the archive an agent can search
The upstream project is MemPalace/mempalace, by Milla Jovovich and Ben Sigman, under the MIT license. It launched in April 2026. Our older article linked the wrong repository; that link has been corrected.
PyPI lists 3.11.0, published October 14, as the current package. The development branch advertises 3.12.0. Those are different things. The published package metadata permits Python 3.9 or newer, while the current installation guide asks for Python 3.10 or newer. Follow the guide's higher floor rather than assuming every allowed interpreter has been tested.
Install mempalace with pip inside a dedicated Python environment. Use mempalace init for project setup and mempalace mine to import selected material. The documentation covers conversation exports, Claude Code sessions and project files. Mining means reading and indexing that material; it does not mean the agent automatically knows every file on your computer.
ChromaDB is the default local storage backend. The released embedding code supports MiniLM, multilingual EmbeddingGemma and an OpenAI compatible embedding endpoint. CPU execution is supported. EmbeddingGemma's documented model download is approximately 300 MB. Allow additional disk for packages, original text, indexes and backups. We found no dependable RAM minimum for a growing archive and have not measured one.
Local embeddings and local storage can work without a paid inference API. The first model download needs a network connection. Remote embeddings send text to that endpoint; optional model based reranking sends candidate passages to its model. A cloud assistant reading local search results also sends those results to its provider.
The palace names are filing aids
The current palace documentation describes:
- Wings for people, projects or broad subjects.
- Rooms for topics inside a wing.
- Halls for categories such as facts, events, discoveries and preferences.
- Closets for compact summaries pointing toward original material.
- Drawers for stored text and its metadata.
What actually gets saved is drawer text plus metadata. Wing and room filters narrow a search, and the current docs say plainly that this is standard metadata filtering rather than a new retrieval method. It is good filing.
Raw mode keeps original wording rather than relying entirely on extracted facts. MemPalace also has a knowledge graph for relationships and an experimental abbreviation system called AAAK. Do not confuse that compressed representation with a lossless archive. The maintainers explicitly corrected that claim.
Agents reach MemPalace through its MCP server. MCP is a shared interface for calling tools. Examples include mempalace_search, mempalace_add_drawer, graph queries and diary operations. Claude Code and Codex connection examples are documented, as are session hooks. Installing a skill description alone does not install the database or connect those tools.
What happened to the benchmark headlines
MemPalace's team initially promoted 96.6% retrieval recall on LongMemEval and 100% with an optional reranking pipeline. Its recall at five means at least one labelled relevant session appears among five retrieved sessions. It does not require finding all evidence or mean the assistant answers correctly 96.6% of the time.
The primary criticism is public. Issue 27, filed by lhl, challenged compression claims. Issue 875, filed by dial481, challenged comparisons between retrieval recall and competitors' answer accuracy. That author disclosed working on a competing memory project. The point stands anyway: retrieval recall and answer accuracy measure different things, so comparing them cannot pick a winner.
The maintainers' correction record acknowledges the problems. They withdrew the lossless compression description and the claimed 34% palace advantage, corrected token counting, and removed misleading comparison tables. Their corrected headline is raw mode's 96.6% retrieval recall. AAAK's reported 84.2% was worse on the same benchmark.
The benchmark methodology says the final step to 100% targeted three known failing questions. It reports 98.4% on a held out set instead. It also explains that a historical LoCoMo result retrieved every session because its cutoff exceeded the number of available sessions. That cannot demonstrate better retrieval discrimination.
Those corrections make the project easier to assess. They do not establish how well it will find your customer's disputed instruction. We have not reproduced these benchmarks.
Can you use both today?
Yes, there are documented and implemented integration routes. Hermes supports local MCP servers, and MemPalace exposes one. Configure that server under Hermes's mcp_servers settings, then verify that its tools appear and return the intended palace's records.
There is also a native Hermes memory provider in MemPalace's v3.11.0 source. It implements Hermes's provider interface, with retrieval and background filing of conversation turns. This is stronger evidence than two product names appearing in a compatibility list, but we have not tested the installation together. A provider must actually be installed and registered; changing a configuration value alone is insufficient.
MCP exposes callable tools. A provider can participate in automatic recall and saving. Choose the behavior you need before enabling both and accidentally keeping duplicate histories. The upstream provider also warns that Hermes's own backup does not include MemPalace's separate data. Back up both.
Which one fits which job
For a phone based assistant that checks a queue each morning and drafts a reply, we would try Hermes Agent first. Its scheduler and messaging gateways address the missing work. Start with a harmless report and require review before anything gets sent or changed.
For a developer who keeps asking why an old migration was cancelled, we would try MemPalace. Import a bounded project archive, ask questions whose answers you already know, and inspect the original passages. If you already use Claude Code or Codex, another agent runtime may add nothing to that task.
For someone who needs scheduled work and a shared archive across different agents, using both makes sense. Keep authoritative instructions separate from retrieved conversation fragments. An old suggestion should not override today's operating rule.
For a team that requires everything to stay inside the building, either can be part of the design, but neither name guarantees it. Keep answering models, embeddings and storage local. Inspect optional providers, gateways and tools for external calls. Restrict network access and test it. A local database attached to a cloud model is still a cloud workflow.
For a CPU only server, MemPalace's local embeddings are a plausible starting point. Hermes Agent can also run there with remote inference. Fully local answering depends on the chosen model, context length and acceptable wait. We have no measured speed comparison to promise.
For stable preferences and a handful of project rules, we would use neither. A short Markdown memory file is easier to inspect, correct and back up. Add retrieval when finding old material becomes the actual problem.
Other options include Mem0 for a memory layer with library, server and hosted routes, Graphiti and Zep for relationships that change over time, and Letta for stateful agents. Honcho is another option for modeling users across conversations. Each one makes a different design bet, so pick by the job you have.
Give memory less authority than the person
A malicious page can plant an instruction that an agent later saves and follows. So can an imported transcript. Memory poisoning persists beyond the original interaction. Keep passwords, access tokens and unnecessary client records out of memory. Scope imports by project, preserve provenance, review important saved decisions and test restoration from backup.
Hermes's security policy describes a personal, single tenant agent. Its default terminal can run commands on the host. Alternatives include Docker, SSH, Modal, Daytona and Vercel Sandbox, but isolating the terminal does not isolate every plugin, hook or MCP process. Whole process isolation matters. Approval prompts and content scanners are helpful checks, not containment.
There have been real fixes. CVE-2026-10222 concerns Hermes configuration injection; the advisory lists 0.18.0 as the patched version. MemPalace's remote MCP advisory documents a server mode intended only for reading that left some write and deletion tools callable, fixed in 3.7.0. Update deliberately and read the notes.
MemPalace is still maintained, with October commits and active issue threads. Its security policy distinguishes its default embedded Chroma client from a separately exposed Chroma HTTP server. Do not turn local storage into a public service casually.
What we run
On our own CPU only Linux server, with 48 cores and 94 GB of RAM and no GPU, we run the Hermes 3 8B model through Ollama. It is the last fallback in our self hosted prompt hub, after Claude Code and Codex. That is a Hermes model, not Hermes Agent.
Our memory today is a nightly semantic index of notes, task briefs and audit deliverables in Qdrant, using a 768 dimension embedding model. The coding agents also keep plain Markdown memory files. Our model routing article explains the separate roles. We did not benchmark any of it for this post.
We have not installed Hermes Agent or MemPalace on that server. We are weighing them now. A useful trial would have to improve retrieval or remove a real recurring chore without making the existing workflow harder to trust.
FAQ
Does Hermes Agent require a Hermes model?
No. It can use supported hosted providers or a compatible local endpoint. Choose a model that handles your tools reliably.
Does MemPalace replace Claude Code or Codex?
No. It supplies stored context and retrieval tools. The coding agent still does the work.
Will either keep everything private automatically?
No. Storage location, answering model, embeddings, messaging and optional services each affect where data goes.
Do I need a GPU?
Neither the agent runtime nor MemPalace's CPU embedding route inherently requires one. Local language model performance is a separate question.
Does a 100% memory benchmark mean perfect answers?
No. Retrieved evidence, generated answers and tests tuned around known questions measure different things.
What should I try first?
Write down the repeated failure. Try a curated note for forgotten rules, retrieval for missing history, or an agent for unattended tasks. Test with material you can safely inspect.
Fact-check notes and sources
All linked primary sources were read on October 29, 2026. Version statements distinguish published releases from development documentation. Benchmark figures are attributed reports, not our measurements.
- Nous Research: release catalog, Agent release notes, installation, memory, provider and security documents linked above. Read October 29, 2026.
- Nous Research: Hermes model cards and the Hermes 3 report linked above, for bases, sizes, licenses and tool formats. Read October 29, 2026.
- MemPalace: package metadata, release history, recent commits and open issues, plus documentation and tagged provider code linked above. Read October 29, 2026.
- MemPalace: correction history, benchmark methodology and community issues linked above. Read October 29, 2026. The older explainer's benchmark comparison should not be read as a current ranking.
- Hermes and MemPalace security advisories, and Mem0, Graphiti and Letta repositories linked above. Read October 29, 2026.
- Our own server setup as of October 29, 2026, also described in the linked prompt hub and routing articles. We ran no installation, latency test or benchmark for this post.
Related reading
- MemPalace explainer and setup generator. Treat the generator's output as a starting worksheet to check against current upstream instructions.
- The case for an agent that remembers.
- How agents coordinate and remember.
- When Markdown memory is enough.
- Where skills, rules and memory belong.
If you set this kind of thing up for clients, The $100 Network covers the business side of running repeatable services. A memory system is one part of that. The service still needs a clear owner and a result the customer can see.
This post is informational. Mentions of Nous Research, Hermes, MemPalace, Anthropic, Claude, OpenAI, Codex, Ollama, Qdrant, ChromaDB, Mem0, Zep, Letta and Honcho are nominative fair use. No affiliation is implied.
