hermes-memra 4.5.0

Memra memory provider for Hermes Agent — EU-native, privacy-first persistent memory with hybrid recall, supersede chains, and staleness signals. Cloud or fully-local.

Memra ↔ Hermes Agent

A Memra memory-provider plugin for Hermes Agent (Nous Research).

Hermes Agent supports pluggable, single-select external memory providers (Honcho, Mem0, Hindsight, OpenViking, Holographic, RetainDB, ByteRover, Supermemory). This makes Memra an option alongside them: hybrid semantic + structured recall, typed memories, importance ranking, EU-native self-hosting, and server-side compression of long-lived memories.

memra/ is the drop-in plugin directory. In the Hermes source tree it lives at plugins/memory/memra/ (the bundled-provider layout, imported as the package plugins.memory.memra). A user install instead goes one level deep at $HERMES_HOME/plugins/memra/ — see below.

Install (pip — recommended)

pip install hermes-memra

Hermes Agent discovers the provider automatically via the hermes_agent.plugins entry point — no files to copy. Then set your key and select the provider:

export MEMRA_API_KEY=memra_live_...   # free key at https://usememra.com
hermes memory provider memra

Fully-local mode (no cloud): pip install memra-local, run memra serve, and set mode: local in the plugin config — same tools, zero external calls.

The install.sh script remains for source-tree / $HERMES_HOME drop-in installs.

What it does

Hermes lifecycle	Memra behavior
prefetch / queue_prefetch	Background hybrid recall before each turn
sync_turn	Persists each turn as an event memory
on_pre_compress	Ships about-to-be-discarded context to Memra so it survives window compression (Memra then compresses it server-side)
on_memory_write	Mirrors Hermes MEMORY.md / USER.md writes into Memra
Tools	memra_search, memra_remember (store or supersede a fact), memra_profile

All network calls are wrapped in a circuit breaker so a Memra outage never blocks the agent loop. The plugin is self-contained (only httpx) — it calls the Memra REST API directly, no Memra client package required.

Install today (Hermes ≥ 0.11)

Hermes discovers user-installed providers one level deep under $HERMES_HOME/plugins/<name>/. The easiest path is the installer:

curl -fsSL https://raw.githubusercontent.com/usememra/hermes-memra/main/install.sh | bash
hermes memory setup        # select "memra", paste API key + project id

Or do it by hand — note the destination is plugins/memra/, not plugins/memory/memra/ (that's the in-tree bundled path):

cp -r memra ~/.hermes/plugins/memra
hermes memory setup        # select "memra", paste API key + project id

Or wire it manually:

hermes config set memory.provider memra
echo "MEMRA_API_KEY=memra_live_xxx" >> ~/.hermes/.env
echo '{"project_id": "proj_xxx"}' > ~/.hermes/memra.json

Get an API key and project id at usememra.com, or point base_url at your own self-hosted Memra to keep all data on your infra.

Want zero cloud? Skip the account entirely and run fully on-device — see Run fully local below. It's a single "mode": "local" switch.

Run fully local (memra-local)

Run everything on your own machine — no account, no API key, no network. The plugin points at memra-local, a single-process on-device memory server (SQLite + on-device embeddings). This is distinct from self-hosting the cloud API above: memra-local runs entirely locally.

Cloud vs local is one switch. The plugin has a mode setting:

mode	Backend	Needs an API key / account?
cloud (default)	Hosted / self-hosted Memra API	Yes
local	memra-local on this machine	No

Set mode to local and the plugin auto-fills the local URL (http://127.0.0.1:8765/v1), a throwaway API key, and a default namespace — nothing else to configure.

What you'll need

• Python 3.11+ with pip (to install and run memra-local).
• Hermes Agent ≥ 0.11 (this plugin).
• ~300 MB disk for memra-local (no PyTorch — it uses an ONNX runtime ~55 MB plus a one-time ~90 MB model download). Your memories live in a local SQLite file.
• A terminal you can leave open (or a background process) for the server — it must be running whenever you use Hermes.

Quick start

Option A — one command (installs the server, configures the plugin, starts it):

curl -fsSL https://raw.githubusercontent.com/usememra/hermes-memra/main/install-local.sh | bash
hermes config set memory.provider memra

That installs memra-local, writes ~/.hermes/memra.json with "mode": "local", and starts the server on port 8765 in the background.

Option B — step by step (you stay in control of each piece):

# 1. Install the on-device server
pip install memra-local

# 2. Start it — LEAVE THIS RUNNING (its own terminal/tab). Port 8765.
memra serve --port 8765

# 3. Install the plugin in local mode (in a second terminal)
MEMRA_MODE=local \
  curl -fsSL https://raw.githubusercontent.com/usememra/hermes-memra/main/install.sh | bash

# 4. Select Memra as the provider
hermes config set memory.provider memra

Either way you end up with a one-line ~/.hermes/memra.json:

{ "mode": "local", "project_id": "hermes", "tenant_id": "hermes-user" }

To switch an existing cloud install to local, just add "mode": "local" to that file (and start the server). To go back, set it to "cloud".

⚠️ The server must be running

This is the one thing that trips people up: memra-local is a separate process that has to be up (listening on http://127.0.0.1:8765). If it isn't, every memory call fails and the plugin's circuit breaker quietly pauses retries for a couple of minutes — looking like "memory just doesn't work."

• Check it's up: curl http://127.0.0.1:8765/health → {"status":"healthy",...}
• It does not auto-start on reboot. After a reboot, run memra serve again.
• Background it: nohup memra serve --port 8765 >~/.memra-local.log 2>&1 &
• Always-on: run memra serve as a systemd --user service.

Good to know

• /v1, not /api/v1 — memra-local mounts the API at /v1. mode: local handles this for you; only matters if you set base_url by hand.
• The API key is ignored — memra-local doesn't authenticate. The plugin still needs some non-empty key, so mode: local sets a placeholder. Don't paste a real memra_live_… key into a local config.
• project_id / tenant_id are local namespaces — any names you like; they scope memories within the on-device store.
• Stale env wins: a MEMRA_BASE_URL / MEMRA_API_KEY left in ~/.hermes/.env overrides mode: local. If local mode seems to hit the cloud, clear those. The init line in ~/.hermes/logs/agent.log prints the resolved mode= and base_url= so you can confirm.

What runs local: the plugin's full REST contract — memra_remember (add and supersede), memra_search, memra_profile, plus the supersession audit chain — is verified end-to-end against memra-local. Recall is semantic (on-device all-MiniLM-L6-v2 embeddings via ONNX, with FTS5 keyword as fallback), so "recall by meaning" works fully offline — no OpenAI key and no PyTorch.

Ship to all Hermes users (PR)

To appear in Hermes's official provider list, the memra/ directory is submitted to NousResearch/hermes-agent under plugins/memory/memra/:

1. Fork NousResearch/hermes-agent.
2. Copy memra/ → plugins/memory/memra/.
3. Add an entry to website/docs/user-guide/features/memory-providers.md.
4. Open a PR. (See docs/developer-guide/memory-provider-plugin.md for their contribution requirements — this plugin already follows that contract.)

PR note for maintainers: unlike the other providers this one ships self-contained (httpx only) rather than depending on a memra-sdk PyPI package, because the import name collides with an unrelated existing memra package. Happy to switch to the SDK once it's published under a non-colliding name.

Test

The unit tests stub out the host imports, so they run standalone — no Hermes checkout and no live Memra account required:

pip install pytest
pytest tests/

They cover tenant-id resolution, the memra_remember supersede paths, and the migrate_tenant cascade handling.

Troubleshooting

Local mode: memory does nothing / every call fails. The memra-local server isn't running. It's a separate process — start it with memra serve --port 8765 and confirm curl http://127.0.0.1:8765/health returns healthy. After a reboot you must start it again. See Run fully local.

Local mode seems to hit the cloud anyway. A leftover MEMRA_BASE_URL or MEMRA_API_KEY in ~/.hermes/.env overrides mode: local. Clear them. The init line in ~/.hermes/logs/agent.log shows the resolved mode= and base_url=.

Very long memories occasionally fail to store with some local models. This is a host-side (Hermes) limitation, not a Memra one. Some local models (e.g. certain OpenRouter local models) emit slightly-malformed JSON when a tool call carries a large argument; Hermes' tool-call sanitizer may drop the argument before it reaches Memra. Memra itself accepts content up to 10,000 characters of any shape. Workarounds: use a well-behaved model (Claude, GPT, most hosted models), or split very large memories into smaller writes. Tracked in #4.

License

MIT © Ali Vonsensey