Skip to content

🎛 Options

Every flag of saltChat, saltServe and salt-mcp, one line each, plus the salt flags a normal compression needs. The concepts live on the Chatbot mode, Serving, Agents, MCP and Architecture pages, and the full detail lives in the code. Defaults shown are what a plain run uses.

saltChat quality switches

Off by default. Each one changes what the conversation memory keeps or resurfaces, and /stats reports what it did, so it can be judged on a real session before it is trusted.

Flag What it does Turn it on when
--conversation-map opens the memory block with a one line per turn index of the conversation the model should see a topic came up even when none of its sentences were selected
--dedup-cos 0.92 skips a new sentence too similar to an earlier one from the same speaker restatements are inflating what counts as a theme
--coverage-half-life 8 fades a shown theme's discount over turns of silence topics the conversation returns to should resurface
--coverage-decay-docs applies that fading to attached files too selection keeps circling a document's head instead of advancing
--shift-damping 0.25 on a detected topic pivot, lifts stale discounts for that turn only pivots back to an old topic are fought by its accumulated discount
--per-source-themes profiles the conversation and each attached file separately a large attachment crowds the conversation out of memory
--query-identifiers lets the question's dates, versions and numbers match memory directly questions that hinge on an identifier keep missing the turn that named it
--episode-gap 6 groups memory into time episodes, splitting where exchanges sit more than this many hours apart, and gives each its own branch a session spanning days blurs its epochs together and answers from the wrong one
--assistant-weight 0.5 counts model authored rows as a fraction of a row in the theme profile the model's longer restatements crowd the user's own topics out of the themes
--row-coverage remembers what was already shown per sentence instead of per tree branch long sessions repeat material because the tree reorganized and dropped its discounts
--stable-coverage-keys freezes the session's keyword order so remembered discounts keep matching their branches long sessions re-show material because themes reshuffled
--coverage-gc collects remembered keys that no longer match any branch of the memory tree long sessions carry dead suppression in every save
--coverage-max-keys 500 hard limit on remembered theme keys you want a strict bound no matter what else is on
--max-sentences 400 keeps the most recent conversation sentences in memory and masks older ones instead of deleting them a long session keeps slowing down as its memory grows
--short-turns fuse stores a bare "yes" together with the question it answers terse decisions should be findable by the question's own words
--memory-floor auto reads at least 400 words of memory, or all of it while the session holds less a young session answers from a few lines of what it remembers
--memory-hold auto reads at most 400 words on a turn the memory holds nothing for a long session spends its whole memory block on requests that stand on their own

saltChat reference

Models:

Flag Default What it does
--model NAME the single registered model, if there is one load a registered model and start the REPL
--add HF_ID none download and register a model
--alias NAME derived short name for --add
--dtype T bfloat16 weight dtype at registration
--force off re-register over an existing alias
--list off list registered models and exit

Session:

Flag Default What it does
--conversation-id ID fresh id resume or name the session
--doc PATH none ingest a text or PDF into the trie at startup, repeatable
--tail N 4 exchanges kept verbatim after compaction
--turns FILE none replay a JSON or JSONL file of user turns into one session
--turns-field KEY auto which key holds the message in --turns items
--turns-out FILE none append each --turns answer as JSONL
--turns-mode conversation\|independent conversation one session for all items, or a fresh session per item so unrelated items never see each other's memory
--turns-timeout SECONDS none give up on a turn whose server sent nothing for this long and go on (served backends)
--turns-resume off keep the --turns-out file and skip the items it already answers, so a stopped run continues
--turns-gold-field KEY auto which key holds the reference answer in --turns items, so each row is scored and the run ends with its count
--turns-template TEXT none compose each --turns message from an item's fields, as in Solve: {puzzle}

Agents:

Flag Default What it does
--roster FILE none JSON file naming the worker models this session may reach, listed by /roster. auto fits one from the registry and the memory free right now, and writes it to the session folder
--personas PATH none load persona role files: a directory of *.md files, one file, or the word samples for the shipped set. Repeatable. Personas ride a roster worker or the chat model itself, so agents work on one GPU or none
--workers-autostart off start the roster's spawn entries once the chat model is loaded, instead of waiting for /worker start
--offload-timeout SECONDS the standard call timeout how long to wait on a quiet worker during a delegation, for workers whose roster entry names no timeout of its own
--offload-budget-pct the session budget memory budget for a delegation's context, as a fraction like --budget-pct
--offload-context-cap N off cap the memory handed to a worker at N words, on top of the memory budget that already sizes it
--offload-ingest-cap CHARS 2000 how much of a worker's answer is remembered, cut at a sentence boundary (0 keeps all of it)
--offload-ingest off remember what a worker answered, as a turn of its own labeled with the worker it came from
--agent-keep-think off keep a worker's reasoning when its answer is remembered, instead of cutting it
--agent off plan every turn out instead of answering it directly, the way /agent does one turn
--agent-quiet off leave out the one-line notice an agent-routed reply carries under --agent
--agent-rounds N 1 how many rounds of delegating one turn may take, at most 2: the second lets the orchestrator ask for one more thing
--agent-think MODE template which parts of a round reason out loud on the models that offer the choice: template, plan, on or off
--agent-max-delegations N 4 how many pieces one /agent turn may hand out before the rest are reported as not attempted
--agent-max-wall SECONDS 600 how long one /agent turn may spend handing pieces out before it answers with what it has
--log-signals off write one line per turn to signals.jsonl in the session folder, holding the numbers this session reports about itself
--route-agent off let a policy decide which turns are planned out over the helpers instead of planning every one of them
--route-rules FILE none the rules --route-agent decides by, as sentences about the session, the ask and the last round
--route-rules-allow-examples off load the route rules a file marks as examples too, none of which has been measured
--route-policy rule\|model rule what decides under --route-agent: the rules file, or the model itself (experimental)
--switch-agent off let a policy decide the memory switches per turn instead of leaving them where the flags set them
--switch-rules FILE none the rules --switch-agent decides by, as sentences about the session and the switch each one changes
--switch-rules-allow-examples off load the rules a file marks as examples too, which are written down to be read rather than run
--switch-policy rule\|model rule what decides under --switch-agent: the rules file, or the chat model itself (experimental)

Memory sizing:

Flag Default What it does
--budget-pct P 0.20 share of the remembered words the memory block may use
--memory-cap N\|auto\|window\|off auto absolute ceiling on the block, fitted to the model's window and never above 4096 tokens
--memory-floor N\|auto\|off off the least a turn reads of memory, in words: a young session is read through at least that many, or all of it when it holds less, auto is 400
--memory-hold N\|auto\|off off the most a turn reads when memory holds nothing for it, in words: a turn the remembered rows do not bear on reads a block of that size in place of the full one, on a memory of 5000 words or more and never on a turn about the memory itself, auto is 400
--branch-stats off show in /stats how each branch of the memory trie, every attached file and the conversation, relates to the last question
--scope off\|auto auto auto searches only the attached files a question is about, plus the conversation, and sizes the block from the words in that scope, off searches every file
--scope-margin M 0.02 under --scope auto, how far below the best file's whole-file similarity a file may sit and stay in scope
--scope-peak-margin M 0.10 under --scope auto, how far below the best file's best-sentence similarity a file may sit and stay in scope
--summary off\|auto off auto profiles more themes and spreads the budget across them on a turn that asks for a summary
--summary-themes P 0.6 under --summary auto, the keyword-frequency percentile a summary turn profiles themes at, lower admits more
--summary-lam L 0.3 under --summary auto, the coverage discount a summary turn selects with, lower spreads the budget across more branches
--gate off\|auto off auto asks the chat model once, with the request alone, whether the turn needs the memory of earlier conversations, and skips the search only when the model is sure the turn needs none (served backend only), off always searches
--when off\|auto auto auto reads a day, month, year or span off the question and holds the conversation rows filed outside it out of that turn's selection, off searches every row

Behavior already on by default:

Flag Default What it does
--short-turns off\|keep\|fuse keep keep short user messages in memory, off restores dropping them
--no-tail-exclude off let the memory block repeat sentences still shown in the recent messages (--tail-exclude is the default and accepted as a no-op)
--no-turn-labels off drop the turn and speaker labels from memory excerpts
--sync-ingest off index messages inline instead of on the background worker

Backends and hardware:

Flag Default What it does
--backend hf\|vllm\|vllm-serve hf how the chat model runs
--server-url URL localhost 8000 the vllm-serve backend's server
--device D auto device for the chat model
--gpu LIST none card index or comma list, several cards shard the model
--bge-device D same as model device for the encoder
--gpu-mem-util F 0.80 across several cards, 0.85 alone fraction of each card the engine may claim
--max-model-len N model's own cap the context window

Tuning values for switches above:

Flag Default What it does
--shift-margin COS 0.12 cosine drop that counts as a topic pivot
--shift-query-boost X 1.5 query boost on a pivot turn while damping is on

saltServe

Launches a persistent vllm serve process for a registered model, so the server and its cache outlive individual chats. Everything after -- passes to vllm serve unchanged.

Flag Default What it does
MODEL the single registered model registered alias or HF id to serve
--host ADDR 127.0.0.1 bind address
--port N 8000 listen port
--gpu LIST none card index or comma list, several cards tensor parallel the model
--gpu-mem-util F 0.80 multi, 0.90 single fraction of each card the server claims
--max-model-len N model's own cap the context window
--vllm-bin PATH alongside saltServe which vllm executable to run

salt

One shot compression of a dataset or a single document.

Flag Default What it does
--data FILE none LongBench style JSONL to compress
--doc PATH none single document instead of --data
--query Q empty question to bias a --doc selection
--output FILE required where the compressed records go
--selector coverage\|legacy coverage CELF coverage selection or the legacy heuristic
--token-budget-pct P 0.20 share of the document kept
--model NAME BGE small encoder
--device D cuda encoder device
--theme-percentile P mode default frequency cutoff for theme keywords
--lam L 0.5 coverage discount between ranking and set cover
--query-mass R 1.0 query weight as a ratio of document mass
--synthetic off paragraph unit adapter for enumerated tasks
--code off line record adapter for code datasets
--fewshot off few-shot mode: for exemplar-shaped contexts, keep a diverse subset of whole exemplars chosen by the selector
--max-samples N all cap the number of records
--verbose off also print the top theme keywords for each record

The tuning knobs are intentionally not documented here: the shared trie thresholds, the legacy selector's pass knobs, and the mode adapters' sub flags. They exist for experiments. salt --help lists every one of them, and their meaning lives with their code in salt/compress.py and salt/engine/dataset_modes.py.

salt-mcp

Serves the same conversation memory to an MCP client, over the tools on the MCP page.

Flag Default What it does
--sessions-dir DIR the saltChat folder where conversations live
--max-open-sessions N 8 how many conversations stay open at once
--max-ingest-chars N 400000 longest text one call may carry
--doc-root DIR any file the server can read read documents by path only from under this folder
--memory-cap N\|auto\|off auto ceiling on the block a memory read returns, in tokens counted at 1.6 a word: auto is 4096, off sizes the block by the budget alone
--read-only off answer reads and refuse every write, leaving every conversation exactly as it was found
--roster FILE none roster of helper models this server may delegate to
--device D cuda when there is one encoder device
--bge-device D same as --device encoder device, winning over --device
--gpu LIST none card index or comma list, with the encoder on the last card
--version off print the salt version this server carries