Memory Shard reads what your box already remembers — its notes and your past conversations — and turns it into a searchable index the assistant can draw on. Point it at folders of your own and it indexes those too. Open it from the Memory Shard icon on the desktop.
Memory Shard needs a Pro or Max ClawBox AI plan. The first step of its setup offers the upgrade if your box is on Free or is not connected yet. See ClawBox AI.

​
Setting it up

1

Open Memory Shard and press Enable

Setup is three short steps from there, and can be re-run later from its settings.
2

Choose folders to read

Point it at folders of your own — notes, documents, exports. PDFs, Word and OpenDocument files are converted to text on the box before they are indexed; nothing is uploaded to do the conversion. This step is optional: your notes and conversations are indexed either way.
3

Pick when it refreshes

Indexing reads every source and re-embeds what changed, so it is best done while you are not using the box. Choose daily or weekly, and a quiet hour.
4

Let it build the first index

The last step is where the index gets its embedding model, and it offers both: ClawBox AI cloud or On this box. Where your plan covers the cloud model it is already selected, so there is nothing to download and the first pass starts straight away. Pick the box instead and it fetches its own embedding model first — about 640 MB, once.

​
Where the index is built

Settings → Embedding model, inside Memory Shard, is a two-way switch — on every edition:
The Memory Shard window: the memory index card marked Cloud and Healthy, with files, chunks, sources, index size, Index now and Full reindex buttons, and the automatic indexing schedule below.

Memory Shard indexing through the ClawBox AI cloud

The cloud is the default wherever your plan covers it. A box that is connected to ClawBox AI on a Pro or Max plan, and whose cloud model answers, indexes there without you choosing anything — so a new box has nothing to download and its first index starts at once. On this box is the opt-in, and once you pick it your choice stands: nothing moves your index back on its own. A box that is not connected, or whose plan does not cover the cloud model, indexes on itself, names the reason on the card, and moves to the cloud by itself — with the full rebuild that move needs — once a plan that covers it is linked. Switching rebuilds the whole index. That is not a precaution — an index belongs to the model that wrote it, so a switch is always followed by a full pass, and the card says so before you press anything. Search keeps working on the old index until the new one is ready. When the cloud option cannot be picked, the card names the reason: the box is not connected to ClawBox AI, the plan does not cover the cloud model, the service did not answer this box (in which case it is retried on its own and the model on the box works meanwhile). Picking the model on the box yourself is not a refusal: the cloud stays selectable, and nothing moves your index back on its own.
The card names the model your index was actually built with, which is not always one of the two buttons: on the OpenClaw edition an embedder pointed at directly in OpenClaw’s own configuration is reported as Cloud and shown under its own model’s name.
Sending documents to be indexed is what the cloud option is: the text of everything you index becomes the body of those requests. If that is not what you want, pick On this box — it is a first-class option on every edition, and nothing about it depends on a subscription once the model is downloaded.On the edition where ClawBox owns the index itself, those are the only two places it will send anything. Every indexing pass and every search checks the address it is about to use and refuses any other: the box’s own model, reached on the device itself, or the ClawBox AI endpoint belonging to this box’s own account. What is stored is which of the two you chose — never an address, and never a copy of a credential. On the OpenClaw edition the choice is written into OpenClaw’s own memory-search configuration as an endpoint and a key, which is why an embedder pointed there by hand shows up above as Cloud under its own model’s name.

​
What you see while it runs

The app shows files, chunks, sources, anything pending and anything that failed, the index size, when the last pass ran and what triggered it, and live progress through the current pass. Index now runs an incremental pass; Full reindex rebuilds everything. Health is one of three words:
  • Healthy — the index matches the configured embedding model and is searchable.
  • Needs attention — usually the index and the model no longer match, which a full reindex fixes. The card says which.
  • Unavailable — the box could not read the index’s state at all.
Turning Memory Shard off in its settings stops the schedule and refuses indexing by hand. What is already indexed stays searchable. Start over switches it off, clears the schedule and runs setup again, keeping your folders, the model and the existing index.

​
How the assistant uses it

  • On the OpenClaw edition the index is the agent’s own, and it searches it as part of answering you. There is nothing extra to do.
  • On the Hermes edition ClawBox builds and owns the index, and the assistant gets a tool for it. Ask in plain words — “what does the lease say about the deposit” — and it searches your documents, quoting the file each passage came from.
What comes back from your own documents is information for the assistant to read, never instructions for it to follow. That rule is written into the tool itself.

​
Where it lives

The index and its state are device state: they survive updates and are wiped by a factory reset, along with the rest of the device’s data. See Filesystem Layout.