Skip to contentdocs.Growther.si
Download C5

Where your data lives

C5 keeps everything on your computer. This page explains what sits where, how to move it, and what C5 will refuse to do with it.

The four kinds of files #

Kind What it holds Where it may live
Configuration your settings, config/c5.yaml, guardrails, the install record your home folder, or a folder your organisation manages
Databases tasks, notes, deliverables, runs, telemetry: four encrypted databases a local disk on this computer only
Secrets and keys the device key that encrypts the databases, the deployment key, your API keys this computer only, or your keystore
Caches downloaded models, the memory-search engine and its index, temporary files a local disk; regenerable, never backed up — see what is safe to delete

By default all four sit under one folder:

text
~/.growther                      macOS and Linux
%USERPROFILE%\.growther          Windows (configuration)
%LOCALAPPDATA%\Growther\C5       Windows (databases and keys on new installs)

New Windows installs keep the databases and keys under %LOCALAPPDATA% on purpose: that folder never travels with a roaming profile, so signing in on another machine cannot carry an encrypted database away from the key that opens it.

Logs are in logs/ under the same folder: errors.log, and on macOS the sign-in service's service.out.log and service.err.log. An install from an older version may also have an errors/ folder of old error files; nothing writes there now, and C5 clears it out the next time it logs an error. They are plain text, outside the encrypted databases, so on macOS and Linux C5 narrows the folder to your own account at every start.

An older logs/categorization-debug.log. Before C5 v327, every failed task categorisation wrote the first 80 characters of the task's goal to this file, in plain text, on every install. C5 no longer does: it writes the file only while the GROWTHER_DEBUG_CATEGORIZATION variable is on (1, true, yes or on), names a goal there only by a fingerprint and its length, and, on macOS and Linux, keeps it readable only by your account; an existing file is narrowed that way at every start, but never deleted. If yours has one from an earlier version, an administrator may delete it: nothing reads it.

The model's replies stay out of the other logs. errors.log never holds a model's reply to a categorisation or to a task's title, only its length, and C5's own output shows one only while GROWTHER_DEBUG_CATEGORIZATION is on, masked and cut to 300 characters. Up to v329, C5's output carried the start of every such reply (in service.out.log on macOS), and errors.log the start of a reply that would not parse or read as an error; an administrator may delete older copies of those files.

See it in Settings #

Open Settings › Storage › Configuration & Data. The card shows:

  • Source: whether this device chose the location (Local) or your organisation did (Managed by your organization). The Local tile also carries a badge naming how the location was decided — Default, Pointer, Environment or Policy. A policy that names only the databases or the key set, and leaves the rest of the home to this device, still counts as managed: the card says which parts your organisation set.
  • Where things live: one row per kind of file, with the path, a badge that says whether it is on this device, on a network folder, or in a synced folder, and a dot that goes amber when something is not where it should be.
  • Actions: change the local folder, verify the layout, roll back a move, or remove a retired copy after a move.

If C5 is running from an older location than the one now configured, a banner says so until you restart.

Move C5 to another local disk #

Use the card's Change local folder button, or the command line:

bash
growther home migrate --to /Volumes/Work/growther

Both run the same procedure:

  1. Validate the new folder: it must be on a local disk, writable, not inside a protected folder, and have room for one and a half times your current data.
  2. Review the plan: what moves, what stays, and what you can prune (old encrypted backup copies can add up to gigabytes).
  3. Confirm. In Settings, type move. At the command line, answer y — or pass --yes, which is required if you are scripting it, because C5 refuses to schedule a relocation from a non-interactive shell without it. C5 then stops, copies everything with a verified checkpoint of every database, checks each copy against your key before anything is switched, and restarts from the new folder.

The old folder is kept as a retired copy with its secrets shredded, so you can Roll back from the card at any time until you choose Remove retired copy.

To see the current layout or check a folder first:

bash
growther home show
growther home probe /Volumes/Work/growther

probe answers for the use you have in mind, and the answers differ — a share that is refused for the databases is perfectly fine for backups:

bash
growther home probe /Volumes/Work/growther --intent=home           # the default
growther home probe //fileserver/c5 --intent=backup
growther home probe //fileserver/c5 --intent=networkConfig
growther home probe /Volumes/Work/growther --intent=location

The move itself takes options worth knowing before you run it:

Option What it does
--dry-run Print the plan and stop. Nothing is scheduled
--yes Skip the confirmation. Required when stdin is not a terminal
--include-imports Carry the imports folder across as well
--include-backups Carry the newest set of database backups across (often the bulk of the size). Older sets stay with the retired copy
--prune-baks Delete the old encrypted database backups (repository.db.encrypted.bak…)
--prune-old Delete the old data folder, data/old
--prune-env Delete the legacy config/.env file

The three --prune-… options name things that are never copied to the new folder, with or without the option. The option marks them in the plan ([x]) to be deleted with the retired copy instead of kept there.

C5 refuses a folder that already holds a file the move would replace, and names the files. Pick an empty folder, or move those aside first. Merging into a folder that already holds unrelated subfolders is fine.

If a move or a roll back is interrupted — the machine loses power part way through, or a rename fails — the next start deals with it, and your data is never left half-moved. What it does depends on how far the job got:

  • Interrupted while copying or verifying, which is most of the elapsed time: C5 discards the partial copy and starts normally from the old folder. Run the move again. It does not resume, because the old folder may have been written to since.
  • Interrupted after the switch: the next start finishes the job.

Either way nothing needs repairing by hand. To abandon a job instead of letting the next start deal with it, run growther home cancel — or growther home cancel --force for one that keeps failing at every start.

What C5 refuses, and why #

Databases are never placed on a network folder, a mapped drive, a roaming profile, or a synced folder such as OneDrive, Dropbox or iCloud Drive. C5 checks the filesystem at every start and refuses to run if the databases are on one of those. The reason is not caution for its own sake: the database engine keeps part of its state in shared memory that only one computer can see, so two computers opening the same folder would corrupt it, and a sync client can copy a half-written file over a good one.

On a locked-down Windows image the check itself can be blocked from running. C5 then decides from what it can see without help — a network path, a OneDrive folder, a synced root — and starts with a warning rather than refusing. A folder that positively looks like any of those is still refused.

If a configured folder is unreachable when C5 starts, it stops with a message naming the folder and who set it, rather than silently starting an empty second copy somewhere else.

Two computers never share one folder. The lock C5 takes on the data folder now records which machine holds it. A second machine pointed at the same folder refuses to start and tells you which machine has it. If a machine has been rebuilt and a stale lock is in the way:

bash
growther lock status
growther lock break --confirm

Backups and network storage #

Network storage is welcome for what belongs there: encrypted backups, exported audit evidence, and the policy file your organisation may publish.

The backup folder under Settings › Storage is for a folder on this computer — it must sit inside your home directory, and a UNC or network path is refused there. Anything off this machine is set under Settings › Enterprise › Enterprise storage, which takes a network folder, an Azure Blob container or an S3 bucket. A network folder is accepted only if an administrator has first added its parent to the allowed enterprise roots on that same screen; with no roots listed, every network target is refused.

Downloaded models #

Three features run small models on this machine: memory search, voice (offline speech to text) and the local classifier. C5 downloads each model once and keeps it in the cache root, a folder of its own that holds nothing you made. Nothing in it is backed up, and every byte in it can be downloaded or rebuilt. Who publishes each model, and under what licence, is in Open-source licences and downloaded models.

Where the cache root is #

By default the cache root is your C5 folder itself (~/.growther, or %USERPROFILE%\.growther on Windows). It moves when one of these says otherwise, the first one that is set winning:

  1. your organisation's policy (the cacheDir location key);
  2. the GROWTHER_CACHE_DIR environment variable (ignored when the policy sets cacheDir);
  3. the home pointer, which growther cache relocate writes (below).

growther cache show prints where the cache root is and what decided that, the home pointer file, and each folder's size and file count.

What is in it #

Folder What it holds Size
qmd/ The memory-search engine (unpacked from the C5 program), its three models under qmd/cache/qmd/models/, its search index, and a README.md listing each model's address, size and sha256. Also qmd/model-view/ (below), with C5's own checked copies of the models about 2.3 GB with the models, plus up to about 2.3 GB more for the checked copies on a disk that cannot clone files
speech/ The offline speech model, ggml-base.en-q5_1.bin about 60 MB
classifier/ The local classifier's model under classifier/models/, and runtime.json, the settings C5 writes for its process about 812 MB
cache/ Model capability data and embeddings small
lib/ Linux only: the OpenSSL 1.1 runtime C5 extracts for its encrypted databases small

The tunnel program Remote Control downloads, cloudflared, is not a cache: it sits in <C5 folder>/config/rc-tunnel/ with the tunnel's credentials, and C5 checks it against the digest this build records every time it starts the tunnel.

The model files, exactly #

Each model is fetched from one fixed commit of its Hugging Face repository and must match this size and sha256 before C5 uses it. Paths are relative to the cache root.

Feature File Bytes sha256
Memory search (embedding) qmd/cache/qmd/models/hf_ggml-org_embeddinggemma-300M-Q8_0.gguf 333,590,944 b5ce9d77a3fc4b3b39ccb5643c36777911cc4eb46a66962eadfa3f5f60490d63
Memory search (reranking) qmd/cache/qmd/models/hf_ggml-org_qwen3-reranker-0.6b-q8_0.gguf 639,153,184 22c9979ce4fbcdc5acdc310c6641c32797eff1aa980b8f7a2db8a8ea23429a48
Memory search (query expansion) qmd/cache/qmd/models/hf_tobil_qmd-query-expansion-1.7B-q4_k_m.gguf 1,282,438,912 000dfb1c06efa6a049e9f64ba921c3740e2454f62abab6fa10e77bd30bb2bcc0
Voice speech/ggml-base.en-q5_1.bin 59,721,011 4baf70dd0d7c4247ba2b81fafd9c01005ac77c2f9ef064e00dcf195d0e2fdd2f
Local classifier classifier/models/decider-0.8b.Q8_0.gguf 811,844,032 2665d08c1052b4e01dabcb08771d25579f6776f7066355e4a0ccc5f74e32d4a2

Two earlier uploads of memory-search models are also accepted when they are already on disk, so an install that downloaded them before is not made to download again: the embedding model at 328,576,992 bytes, and the query-expansion model at 1,107,408,608 bytes. C5 only ever downloads the versions in the table.

The other files you may see #

File What it is
<model>.partial A download in progress, or one that was interrupted. Memory search and the local classifier resume from it; the speech model's download starts over (it is small) and removes its .partial if it fails. It never has the model's name, so nothing loads it
<model>.partial.lock Held while one C5 process downloads that model, so two processes — a running C5 and a pull in a terminal — never write the same file. A lock left behind by a process that has gone (a crash, a reboot) is recognised as stale and taken over
<model>.verified.json The result of checking the model: its sha256, with the size and modification time it had. C5 hashes each model once, not at every start, and checks again if the file changes
<model>.rejected, <model>.rejected.1, … Memory-search and voice only: a file that was at the model's name but is not the model this build records. C5 renamed it aside so it cannot be loaded. It is never deleted — it may be the only copy you have — so remove it yourself once you no longer need it
<model>.ipull Memory search only: left by an older version's download. C5 installs it if it holds the whole model, and otherwise removes it once it is of no further use
qmd/model-view/<number>/ Memory search only: the folder the running search engine finds its models in, holding only the model files C5 has checked — as C5's own checked copy, a copy-on-write clone, a hard link or, failing those, a symbolic link — so the engine cannot load a file C5 has not checked. C5 removes it when the engine stops and removes any left behind at its next start. growther cache show does not count it and growther cache relocate does not move it. Safe to delete while C5 is stopped
qmd/model-view/checked/<sha256>.gguf Memory search only: C5's own checked copy of each model, made in the background after the engine starts and checked by fingerprint before use, so a file later copied over the model's name cannot reach the engine. On a disk that can clone files it takes no extra space; otherwise each copy takes the model's size again (up to about 2.3 GB for all three), and is made only if 1 GB would still be free. Kept from one start to the next; copies no model needs are removed. Not counted by growther cache show, and not moved by growther cache relocate (C5 makes them again at the new location). Safe to delete while C5 is stopped: C5 makes them again

A file set aside is reported in the log, for example:

text
[mcp] QMD: <path> is not the embedding model this build records (<why>). C5 set it aside as <name>.rejected so QMD cannot load it — nothing was deleted. Where downloads are allowed C5 fetches the pinned file in its place; otherwise place it as <cache root>/qmd/README.md describes. Remove <path>.rejected once you no longer need it.

A memory-search or voice file that changed within the last minute is not judged yet — it may be a copy still arriving — and is not used meanwhile. Memory search checks it again once a minute has passed since it last changed; the speech model is checked again the next time it is used or C5 starts.

The local classifier is handled differently. A wrong file at its name is not renamed: it is left where it is, never loaded, and replaced by the next download that passes its check. Settings › Classifier says The file at the model's path is not the model (…), so C5 does not use it., and growther classifier status reports it (is not the pinned model … growther classifier pull replaces it). A symbolic link at its name is never downloaded over or replaced, whatever it points to: one whose file cannot be reached (a share that is not mounted) is used once it can be, and one that points to a file that is not the model stays unused until you point it at the model or remove it.

Place a model by hand #

This is how a machine without internet access gets its models, and it works the same everywhere:

  1. On a connected machine, download the file. The exact addresses are in <cache root>/qmd/README.md for memory search, and are printed by growther voice pull and growther classifier pull when this machine is offline. Each is https://huggingface.co/<repository>/resolve/<commit>/<file>.
  2. Check its sha256 against the table above.
  3. Copy it to the path in the table, under exactly that name. A file under any other name is ignored. The memory-search files must be renamed: Hugging Face serves them as embeddinggemma-300M-Q8_0.gguf, qwen3-reranker-0.6b-q8_0.gguf and qmd-query-expansion-1.7B-q4_k_m.gguf; save each under the hf_… name in the table.
  4. Let C5 find it:
    • Memory search: while C5 runs, about half a minute after the copy stops changing; C5 then restarts the search engine with it, with no restart of C5 needed. If C5 is not running, at its next start. growther qmd-run pull checks the files already there straight away, before it downloads anything. The search engine uses a placed file only once C5 has checked it.
    • Voice: the next time someone uses the offline speech engine.
    • Local classifier: while it is on, the next time Settings › Classifier is opened, or at the next start, with no restart needed. While it is off, when it is next turned on.

For memory search and voice, C5 never downloads over a file that is already at a model's name: a wrong file is set aside first and, where downloads are allowed, the right one is downloaded in its place. For the local classifier, a download that passes its check replaces whatever wrong file was at the name — but never a symbolic link.

What is safe to delete #

Everything in the cache root can come back. What matters is whether it comes back on its own, and how big that is.

To free Do this What happens next
The local classifier's model (812 MB) As an administrator, Delete the model in Settings › Classifier. Or stop C5 and delete classifier/ by hand Turned off first, it stays gone. If it is still on, C5 downloads it again at the next start, unless the machine is offline or opted out
The memory-search models (about 2.3 GB) Stop C5 and delete qmd/cache/qmd/models/ C5 downloads them again at the next start, unless the egress posture is offline or GROWTHER_QMD_NO_AUTO_DOWNLOAD=1. Until they are back, memory search works by keywords only
The whole qmd/ folder Avoid this unless you mean to start memory search over: stop C5 and delete it The engine is unpacked again from the C5 program and the models are downloaded again as above. It also deletes memory search's index of your folders, which does not come back by itself: memory search finds nothing in them until an administrator presses Sync Now on the QMD Data Folders card under Settings › Locations. Delete only qmd/cache/qmd/models/ if model space is all you want back
The offline speech model (60 MB) Stop C5 and delete speech/ Downloaded again at the next start while the built-in speech engine is on, unless offline or GROWTHER_VOICE_NO_AUTO_DOWNLOAD=1
A .rejected file Delete it Nothing; C5 never uses it
A .partial file Stop C5 and delete it The next download starts from the beginning

Stop C5 before deleting by hand: a running C5 holds models open, and Windows refuses to delete an open file. To keep a model from coming back, use the switch for its feature, the opt-out variable, or the offline posture — see Network and egress.

Move the caches to another disk #

The caches can live somewhere other than the rest of C5's data, for example on a larger local disk:

bash
growther cache show
growther cache relocate --to /Volumes/Big/growther-cache --dry-run
growther cache relocate --to /Volumes/Big/growther-cache

Stop C5 first. relocate copies all five folders, checks each copy file for file, by count and size, and only then switches C5 over and removes the old copies; if it is interrupted, the old cache root is still the one in use. --dry-run shows the plan and stops. --yes skips the confirmation, and is required when the command is not run from an interactive terminal (a script, or a session with no terminal).

It refuses, and says why, when:

  • C5 is running — run growther stop first, then relocate, then start C5 again;
  • GROWTHER_CACHE_DIR is set — that variable outranks the move, so change or unset it instead;
  • the cache root is already the target;
  • the target is inside the current cache root, or contains it;
  • the target is a network folder, a synced folder or another place that is not a local disk — the caches hold programs C5 loads and models it reads directly from disk;
  • the target is not writable, or a folder there already holds files (move or remove them first);
  • a growther classifier pull, growther qmd-run pull or growther voice pull is downloading a model in a terminal at that moment. Let it finish, or stop it, then move.

Check the layout from the command line #

bash
growther doctor
growther doctor --json

The report includes Home source, Data dir filesystem, Secrets location and Lock host. The --json form also writes doctor-status.json into the C5 folder for device management tools to read.

Edit this page on GitHub ↗Last updated Nov 20, 2026