📄 Changelog
What each version added. The milestone versions, the x.y.0 headings,
match the git tags on the
repository. The patch versions
listed under them are not tagged.
3.0.0 - 3.0.49
The agents release. saltChat plans a turn out over helper models, or over personas riding its own chat model when there is one GPU or none, grades the reply it wrote against the material it was written from, re-targets a piece whose helper vanished, and lets a rules file decide which turns are worth planning at all. Its memory now searches where the question points: the attached files it is about, the time it names, and never past a ceiling of 4096 tokens, each of those decisions measured against handing the right material over by hand. A scripted run drives a whole dataset through the same path, with a fresh session per item, scoring, templating and resumption. Everything in the 2.11 line is here unchanged. The agents, chatbot and architecture pages cover the details, and the 2.11 section below lists each step.
- 3.0.8 Switch values are range-checked. A rules file or a model
proposal sets a switch only to a value its flag accepts, so
nullfor the shift margin ortruefor the key limit is refused when the file loads, and dropped with the reason kept when the model proposes it, instead of failing or erasing memory partway through a turn. - 3.0.9 A handed-out piece selects under every memory switch. The
pieces of a planned turn and an
/offloadnow carry--query-identifiers,--episode-gap,--assistant-weightand--row-coveragethe way the turn itself does, instead of selecting without them. - 3.0.10 An install that skips the requirements file stays on
releases SALT runs with. pypdf stops below 6.16, with which the PDF
reader no longer keeps a numbered algorithm together. The MCP extra
asks for mcp 2.0, the first release with the server class
salt-mcpbuilds on, and stops below 2.1, which reports every refused tool call as a bare error. - 3.0.13
saltServe --vllm-binfinds the tools of the environment it points at. That environment's bin directory now comes first on the server's PATH, so a vLLM release that builds kernels as it starts finds its ownninjainstead of failing to launch. Without--vllm-binthe server's PATH is unchanged. - 3.0.14
saltServeasks vLLM for prefix caching on every launch. Some vLLM releases leave it off for hybrid attention models such as Qwen3.5, which cost a served chat all of its reuse from one turn to the next. Ending the command with-- --no-enable-prefix-cachingstill turns it off. - 3.0.16 A helper's server is trusted to hold replies to a schema
only when it shows it. Newer vLLM releases answer a request in the
old
guided_jsonspelling and ignore the schema, so schema-bound agent calls went out unchecked. The probe now asks for a schema that allows one reply, first asguided_jsonand then asstructured_outputs, and keeps the spelling whose reply comes back held to it. - 3.0.17 saltChat loads the tokenizer a model was trained with. When the installed transformers cannot read a tokenizer saved by a newer release, or builds one that splits text differently from the model's own tokenizer.json, as transformers 5 does for some models, every backend falls back to that tokenizer.json. Tokenizers that already load faithfully are untouched.
- 3.0.18 A reply that runs out of room while still reasoning is no longer remembered as the answer. Some chat templates, Qwen3.5-9B's among them, open the reasoning block in the prompt, so a reply cut off before it closed had no tag at all and was stored whole. Such a reply now stays out of memory and the tail, and one line names the setting to change. Models whose template leaves the block to the model are unchanged.
- 3.0.19 A chat model that cannot load ends the start with one line
naming the error instead of a traceback, and leaves no empty session
behind. When the model's config shows why, a second line says so: a
model type the installed transformers does not know, a checkpoint
only the vLLM backend can run, or a vLLM release that cannot run next
to the installed transformers.
/modelkeeps working after a switch where neither model loads. - 3.0.20
--addand/addread the model's ownconfig.jsonbefore the download and warn when this environment cannot load that model, then register it anyway, so a multi-gigabyte download no longer ends in a model that fails only at its first start. - 3.0.21
eval.py --backend vllmstarts on a cold vLLM model cache when numpy is built on Intel MKL, as in the Anaconda defaults channel. vLLM's model check runs in a child process that aborted on an MKL threading clash there, so the first evaluation of a model on a fresh machine failed. A value the user sets forMKL_THREADING_LAYERstill wins. - 3.0.22 Replies on the HF backend stop at the chat's end-of-turn token. A model whose generation settings name only its end-of-text token, Qwen3.5 among them, ran past the end of its turn and left a trailing newline on every reply. Models whose settings already name that token are unchanged.
- 3.0.23 The HF backend names the weights' precision the way the installed transformers expects, so transformers 4.56 and later no longer print a deprecation notice on every model load.
- 3.0.24
salt-mcpkeeps request log lines from the Hugging Face client off its stderr. Newerhuggingface_hubreleases log every file check at startup, which buried the server's own messages. Warnings still show. - 3.0.25 The vLLM backends show a reply the way the model wrote it.
When saltChat's tokenizer loader has to replace the tokenizer the
installed transformers builds, the in-process engine and
saltServenow get the corrected one too, so a model such as DeepSeek-R1-Distill-Llama-8B no longer answers withÄin place of every space under transformers 5. Models whose tokenizer loads faithfully start with the same engine settings as before. - 3.0.26 SALT installs next to transformers 4.55.2 through 5.5, so an environment can load model families newer than the 4.55 line, such as Qwen3.5, on the HF backend. The vLLM extra keeps vLLM 0.11.0 on transformers 4.55.2, since that vLLM release cannot run with transformers 5.
- 3.0.27 The supported stack moves to vLLM 0.19.1, torch 2.10 and
transformers 5.5.3, which load model families such as Qwen3.5 on every
backend. It needs CUDA 12.8 and an NVIDIA driver 570 or newer, and it
still runs on Volta cards. The vLLM backend now installs through the
package's extra,
pip install -e ".[vllm]". The published results page names the previous stack, which produced its numbers. - 3.0.29 The note about a reply that ended inside its reasoning also prints when the model opened the reasoning block itself, as the Qwen3 family does, instead of the answer vanishing without a word. It suggests turning thinking off only when the model's template has that switch.
- 3.0.33
salt-mcpfinishes writing its conversations when a Ctrl-C or kill arrives while it is already writing them on the way out, as a client that waits a moment and then sends one can do, instead of stopping partway through. A second one still stops the server at once. - 3.0.35 - 3.0.37
--gate autolets a turn skip the memory search. The chat model is asked once, with the request alone, whether the turn needs the memory of earlier conversations, and a Yes at a probability of 0.999 or more skips the search for that turn. Off by default, served backend only,/gatesets it per session and/statscounts the decisions. - 3.0.41 A script in the results folder reproduces the legacy
selector's LongBench table for Llama or Ministral, one model per run,
and the folder's page shows the table.
eval.py --tokenizer-modepicks the tokenizer the vLLM backend loads. Newer vLLM loads Ministral with Mistral's own tokenizer, which reads the chat template's markers as plain text and leaves many answers empty, andhfkeeps it on its Hugging Face tokenizer. The defaultautoleaves the backend as it was. - 3.0.43 - 3.0.44 A memory floor.
--memory-floor automakes a turn read at least 400 words of memory, or all of it while the session holds less. The block is a fifth of what is remembered, and a fifth of a young session was a few lines. The defaultoffsizes the block as before. - 3.0.45 A memory read over MCP has a ceiling.
session_memorystops at the 4096 tokens a chat turn stops at, counted as 2560 words, where it returned a fifth of the conversation however long that was.salt-mcp --memory-capmoves the ceiling or turns it off, and the reply says when a block was cut to it. - 3.0.46 - 3.0.48 A memory hold.
--memory-hold autoreads at most 400 words on a turn the session's memory holds nothing for, such as a standalone request in the middle of a long conversation, where the block was a fifth of everything remembered whatever the question. It acts on a memory of 5000 words or more and never on a turn about the memory itself. The defaultoffreads the full block as before. - 3.0.49 The floor and the hold are switches a turn decides. A
switch rule or the model policy sets
memory_floorormemory_holdfor the turns it applies to, held to a whole number of words above zero. The pieces of a planned turn and an/offloadselect under the same two numbers, where they selected without them.
2.11.0 - 2.11.123
The MCP milestone. SALT is a working memory layer for any MCP client:
salt-mcp serves compression, conversations and delegation over the
Model Context Protocol, and behind it saltChat can plan a turn with a
reasoning model, fan the pieces out to helper models in parallel, and
remember what a helper answered under that helper's own name. Every
agent behavior ships off by default and is bounded and recorded when
switched on. The MCP server and agents pages
cover the tools, the roster and the limits.
- 2.11.1 - 2.11.3
orphan_share. The conversation snapshot reports orphaned suppression as a share of the whole coverage table beside the raw mass, so a rule about stale coverage means the same thing in a short conversation and a long one, and the sample rules file's orphan example now reads that share. - 2.11.5 - 2.11.6
--doc-root. The MCP server can be started with one folder it reads documents from, andsalt_ingest_documentrefuses a path outside it. Off by default, so a server started as before still reads any file it can. Worth setting when the client driving the server is a model rather than a person. - 2.11.8
--agent-roundsis checked at launch. A number of rounds no turn can run is refused before the model loads, rather than ending the first planned turn that reaches it. - 2.11.9
--agentreaches a scripted run. A plain line of a--turnsfile is planned out over the roster's helpers the way a typed line is, and its--turns-outrow says whether that turn went through a round or was answered plainly because no worker was ready. - 2.11.10 A turn's memory decision reaches the pieces it hands
out. Under
--switch-agent, the subtasks of a planned turn select their context the way that turn selected its own, instead of under the settings the session was launched with. A delegation that belongs to no turn is unchanged. - 2.11.11 A planning call gets room to reason. A model that thinks out loud used to plan under a chat reply's length, spend it on the working and be cut off before the answer, which reached the round as a reply that was not a plan at all. The allowance is bounded by a quarter of the model's window, and a roster entry that names its own reply length still keeps it.
- 2.11.12 A round's record says what writing it up cost, or says nothing. The figure used to be read off the session's own model whatever had happened, so it was empty on every round that worked and filled in only on the ones that gave up on their helpers and answered directly.
- 2.11.13
--log-signals. A session can write down the numbers it reports about itself, one line per turn insignals.jsonl, which is the same closed set a memory decision is allowed to read. Off by default. Worth turning on for a session or two before trusting a rule whose threshold you have not watched fire. - 2.11.14 Chat template settings can be set. A few models expose a
choice through their chat template rather than through sampling, and
chat_template_kwargsnow reaches the template on every backend. It is rendered locally, so it never reaches the server as a request field, and setting nothing renders the prompt it always did. - 2.11.16 A roster entry can ask a model to reason, or not to, with
"think": trueor"think": false. Leaving the key out is the default and sends no thinking setting at all. The key is a request rather than a promise: a model whose template has no such setting is unaffected, and one that always reasons keeps reasoning. - 2.11.18
--agent-think. A round is three kinds of call, and this says which of them should reason out loud on the models that offer the choice.planasks for reasoning where the turn is decided and for none where its time goes,onandoffask for all three or none, and the default asks for nothing. A roster entry's ownthinksetting still wins. - 2.11.19 A model that never stops reasoning ends its own call. One that has spent three quarters of its reply length on the working without answering is given up on there, keeping what arrived and leaving the worker usable, rather than generating to its cap and reaching the round as a reply that was never an answer.
- 2.11.20 The session's own model can be held to the plan's shape.
Under
--backend vllm-servethe chat model is reached over HTTP like any helper, so its server is asked once whether it will hold a model to a schema and is handed one when it will, which is the difference between a plan that parses first time and one that needs a repair. A model loaded in the session itself is unchanged: it has no request body for a schema to ride on and is shown the worked example. - 2.11.21 - 2.11.25
--route-agent. Under--agentevery line was planned out over the helpers, so a turn that said thanks cost the same three model calls a turn that needed the work did. A rules file can now decide, turn by turn, whether to plan at all and what the plan may spend, written in the same language the switch rules are over a wider set of signals: the conversation, the ask, who is ready to help, and what the last round did. A decision may only spend less than the flags already allow./statsprints how often each rule actually fired beside what its author expected, since a threshold that never fires is the way a decision layer quietly does nothing. Off by default, and every rule in the shipped sample is an example. - 2.11.27
/statscounts a switch rule over the whole session, not only the last turn. Every rule is listed with how many of the turns it was asked about it fired on and the note its author left, so a rule that never fires and one that fires on nearly every turn are both visible after one conversation. - 2.11.29 Room to reason. A helper that writes its working before its answer used to plan under a chat reply's length and spend all of it on the working, so the round got a reply that was never an answer. A roster call that names no length is now sized from the window its endpoint actually serves, and an entry that asks for the working with a reply shorter than 2048 is refused when the roster loads, with the fix in the message.
- 2.11.30 The write-up reads the turn's memory. The model that writes a planned turn's reply was sent the helpers' pieces alone, so it answered without the memory the plan itself was shown. That block now travels with the write-up, cut to whole sections when the pieces leave less room than it needs.
- 2.11.31 A planned turn records what its own calls cost. The trace
line said what every piece cost and nothing about the plan, the
follow-up and the write-up, and its write-up field was filled in only
on turns that gave up and answered directly. Both now land on that
line as
planningandsynthesis. - 2.11.32 A route rule that would freeze does not load. A rule that reads what routing itself decides, without saying how stale a number it will act on, fires once and then answers every later turn from that one reading, because the number only moves when a round runs. Such a rule is refused when the file loads, naming the signals and the fix.
- 2.11.33
--route-policy model. The question of whether a turn is worth planning can go to the model instead of a rules file, which is what a session with no rules to give needed. It is shown the conversation as numbers, what the session allows and who is ready, and what it asks for is clamped against the same flags a rule is. Off by default and experimental. - 2.11.34
--route-agentwithout--agentprints the recipe. Such a session built the policy, read the rules file and routed nothing, saying so nowhere. It now says what the whole recipe is and carries on with routing off, rather than refusing to start. - 2.11.35 - 2.11.36
--roster auto. A roster can be fitted instead of written: SALT reads the models you have registered and the memory free on each card, sizes a helper for each job it knows about, and writes the file to the session folder. A share covers weights, KV cache and a flat reserve with room on top, because a share sized to exactly what a model needs is one vLLM refuses to start on. Each fitted helper is described in words the other one does not use, since a planner shown two identical descriptions has no reason to send one piece here and the other there. When two helpers do not fit it prints the roster it would have written, with the arithmetic for each entry that missed, and writes nothing. Fitting starts nothing on its own. - 2.11.44 - 2.11.46 EMNLP 2026. SALT: Salience-Aware Lexical Trie for Long-Context Compression is accepted to the 2026 Conference on Empirical Methods in Natural Language Processing as a Main Conference paper, held in Budapest. The new Paper page carries the abstract, the entry to cite, and how the selector the paper describes relates to the coverage selector releases default to today.
- 2.11.54 - 2.11.55
--query-identifiers. New off by default switch that lets the question's identifier shaped tokens, like dates, versions and numbers, match conversation memory directly. The letters only keyword gate dropped them before, so a question that hinges on a version number could miss the turn that named it even though memory indexes such tokens./statscounts the terms each turn adds. - 2.11.56 - 2.11.58
--episode-gap. New off by default switch that splits conversation memory into time episodes where exchanges sit more than the given hours apart, and gives each episode its own branch of the memory tree, so selection spreads across time the way it already spreads across attached files. A--turnsitem can now carry atimestamp(epoch seconds or ISO 8601), so a replayed conversation keeps its real gaps instead of collapsing into the minutes of the replay./statscounts the episodes. - 2.11.59 - 2.11.60
--assistant-weight. New off by default switch that counts assistant and worker rows as a fraction of a row in the theme profile. The model usually writes several times more text than the user, so its restatements can author most of the themes. Down weighted, themes track what the user talked about while every sentence stays in memory and stays selectable./statscounts the rows down-weighted each turn. - 2.11.61 - 2.11.62
--row-coverage. New off by default switch that remembers what was already shown per sentence instead of per branch of the memory tree. The tree reorganizes as a session grows, and a discount tied to a vanished branch silently disappears, so long sessions can repeat material. Tied to the sentence, the discount follows it wherever the tree puts it next./statscounts the rows carrying discount. - 2.11.63 - 2.11.69
--personas. A persona is a markdown role file that rides a roster worker or the session's own chat model, so one GPU, or none, is enough to plan a turn across an explainer and a coder with no roster at all. Handing pieces to several servers is a latency optimization: the decomposition, the memory selected for each piece alone and the role each piece is answered in all survive running them one at a time. A refused--roster autotherefore keeps the agent layer instead of ending the launch, printing the arithmetic it refused on. Three sample roles ship,/rosterlists what loaded, and temperature 0 now means greedy decoding so planning works against a locally loaded model. The agents page covers the format. - 2.11.70 - 2.11.71 The checker. A
verifypersona grades every planned turn's written reply against the material it was written from and reportssupported,partialorunsupportedwith the issues it can pin to an excerpt. The verdict is printed and kept in the trace, the reply is never touched, and a checker that fails recordsskippedand costs the turn nothing. - 2.11.72 - 2.11.77 Planned turns that hold together on a real
machine. A piece whose helper went quiet, died, or could not be
reached at all is re-run once on a different set of weights, paid
out of the same turn allowance and carrying the first attempt's fate
on the answer, while a turn with nothing left retries nothing.
Launch warns when two helpers are described in shared words, since
the planner picks them by their notes, and the roles that ship no
longer trip that warning against a fitted roster. Schema-held calls
follow newer vLLM servers to
structured_outputs, so planning and checking stay schema-guided across versions. - 2.11.86 A memory ceiling in tokens.
--memory-cap auto, the default, still fits the block to the model's window and now also stops at 4096 tokens, so a long session or a huge document no longer fills the window with memory.--memory-cap windowkeeps the window fit alone. - 2.11.87 - 2.11.95 Scoped search. A session reports how each
attached file and the conversation relate to a question, whole-file
and best-sentence similarity with keyword and name hits, shown by
--branch-stats, and--scope autosearches only the files a question is about plus the conversation, sizing the block from the words in that scope with a floor for short files,/scopenames the files by hand, and/statskeeps a census of the session's routing. A session without attachments selects exactly as before. - 2.11.97
--scope autoby default. A session with several attached files now searches, for each question, only the files the question is about plus the conversation.--scope offsearches every file as before. - 2.11.109 - 2.11.115 Scripted runs that scale.
--turns-mode independentgives each item of a scripted run a fresh session of its own so no item sees another's memory, with each output row naming the session it ran in, while the default,conversation, is the one session the runs always used. Every turn now reports its time and an estimate of what is left,--turns-timeoutabandons a turn whose server went quiet and moves on, and--turns-resumecontinues a stopped run from its output file, retrying only the items that failed. Each output row now also carries the turn's seconds, the prompt tokens and the engine's cache counts, a thinking model's reasoning underthinkbeside the answer it gave, and the item's other fields. An item that carries a reference answer is scored, under the field--turns-gold-fieldnames or a common one: the row gainsgold,correct, the rule that decided it and a token F1, and the run ends with its count overall and per category.--turns-templatecomposes each message from an item's fields, so a raw dataset runs without a preprocessing pass. - 2.11.117 - 2.11.118 Summary turns. The compressor takes a theme
percentile and a coverage discount of its own for one call, and
--summary autouses that on a turn that asks for a summary: more keywords count as themes and a filling branch gives up its sentences sooner, so the block spreads across more topics. Every other turn selects as before./summary nextmarks a turn by hand, and/statsreports what a summary turn selected under. - 2.11.120 - 2.11.121 Time windows. A question about what was said
on a given day was answered from the whole history.
--when autoreads a day, a month, a year or a span such aslast weekoff the question and holds the conversation rows filed outside it out of that turn's selection, files never, so a day that fits the budget is handed over whole./when <time>pins a window by hand, and/statsand the ledger report the window a turn searched. - 2.11.123
--when autoby default. A question that names a day, a month, a year or a span now searches only the conversation rows filed inside it, attached files untouched.--when offsearches every row as before.
2.10.0 - 2.10.123
The agent line. saltChat is growing an agent layer: a session can name smaller helper models beside the chat model and, over the 2.10.z releases, hand parts of its work to them with SALT memory in the loop.
Patch releases, grouped where several versions shipped one thing:
- 2.10.0 Model roster. A
salt.agentspackage that reads a validated roster file naming helper models and how to reach them, with a sample atsalt/agents/roster_sample.json. Groundwork only, nothing in the chat changes yet. - 2.10.1 - 2.10.9 Workers in a session.
--roster FILEloads the roster at launch,/rosterand/workerlist what is there and probe it for the model each endpoint is actually serving, and/statsnames the workers. A session with a roster loaded chats exactly as one without. - 2.10.10 - 2.10.17 Starting and stopping workers.
/worker startlaunches asaltServeworker as a child of the session,/worker stoptakes it down, and--workers-autostartstarts the roster with the session. A worker is refused onto a card another server already claims unless both sides declare their share. A worker that goes quiet mid-reply is given up on, a refused connection is retried once, and one that fails twice in a row is left alone until a probe revives it. - 2.10.18 - 2.10.26 Offload.
/offload <task>hands one task to a worker together with this conversation's memory, selected for the task the way a chat turn selects it and committing nothing, so the session is the same after a delegation as before it. Every delegation is recorded indelegations.jsonl. With--offload-ingestthe answer is remembered as a turn of its own, headed with the worker it came from rather than shown as something you or the model said. - 2.10.27 - 2.10.34 Bounds, resume and scripted delegations.
--offload-context-cap,--offload-budget-pctand--offload-timeoutbound what a delegation is handed, selects under and waits for./statsreports per-worker totals, and a reopened session carries its numbering and totals on while retiring a worker record no live process backs. A--turnsfile can carry an{"offload": ...}item, so a scripted conversation delegates where it needs to. - 2.10.35 - 2.10.41 Delegating with fewer keystrokes. TAB completes
@NAMEfrom the roster,/offload! @NAMEputs the last task to a second worker so two models answer one question from the same memory, and asking with no roster loaded prints the recipe for having one. A mixed chat and delegation conversation ships assalt/agents/demo_turns.json, and the sample roster now shows both an attach and a spawn entry. A delegation runs on the session's own thread, and interrupting one twice still leaves its record behind. - 2.10.42 A worker can answer a turn. A line typed as
@NAME questionsends that turn's own prompt to a worker and keeps the answer as the session's own, remembered and recorded like any other, stamped with the model that gave it. This is changing model in the middle of a conversation without ending it, including under--backend vllm-serve, where the server holds one model and/modelcannot switch it. - 2.10.44 - 2.10.52 MCP server.
pip install -e ".[mcp]"adds asalt-mcpcommand that speaks the Model Context Protocol over stdio, so an editor or an agent runtime reaches SALT directly.salt_compresscompresses one text under a budget, optionally biased to a query. Conversations are the other half:session_create,session_resume,session_listandsession_statsopen and report the same sessions saltChat keeps,session_add_turnremembers a message or a whole exchange in one call,session_memoryreturns the labeled memory block a chat turn would be given, andsalt_ingest_documentreads a file or a text into a conversation under its own source name. Open sessions are held warm up to a cap, and the one used longest ago is written to disk before it is closed. A server started with--read-onlyanswers every read and refuses every write, leaving the conversations it was pointed at exactly as it found them. The MCP server page covers the install, the client entry and every tool. - 2.10.54 Helper models over MCP. Started with
--roster FILE, the server can reach the smaller models that file names:roster_listreturns what is in the roster and, on request, which endpoints are actually answering, andsalt_delegatehands one task over with a conversation's memory selected for it. Selecting it commits nothing, so a conversation is the same after a delegation as before it, and every delegation is filed beside the conversation it ran under. - 2.10.55 A conversation as numbers.
session_statsnow carries a snapshot block describing the conversation in one flat set of signals, from how much of it is still selectable to what its last read measured, andsalt_switcheslists the memory switches, what the server has each one set to, and which measured number says whether it did anything. - 2.10.56 Refusals a client can act on. Every refused call now opens
with a fixed phrase naming what kind of refusal it is, from a bad
argument to a conversation nobody made, and an unexpected fault is
reported as one rather than as a traceback over the wire.
--max-ingest-charsbounds how long a text one call may carry. - 2.10.57 Closing down safely. A conversation closed to make room
finishes what it was still encoding before it is let go, a
conversation whose files disagree after a crash opens repaired and
says so in a
warningslist on the reply, and a server told to stop writes every open conversation before it goes. - 2.10.59 A contract a client can check.
salt_contractreports which version of the tool surface the server speaks, alongside the SALT version and every tool in order. The surface is written down in the server itself, so a renamed or dropped tool stops it at startup instead of surfacing as a broken client. - 2.10.64 - 2.10.67 Planning in a form a session can act on. A model
asked to decide what to delegate answers with one JSON object, read
through the prose, fences and reasoning a local model wraps it in and
held strictly to its shape inside. A reply that is not one is
repaired once, quoting the actual fault, and a second failure keeps
the model's own words as the answer instead of ending the round. A
worker is asked once whether its server accepts a schema at all,
since a version string does not answer that. Reasoning between
<think>tags is cut before an answer is remembered, and--agent-keep-thinkkeeps it. - 2.10.68 - 2.10.71 Knowing what a helper can be asked for.
/roster probe --deep NAMEasks one helper to return three small objects exactly as given and reports whether it is schema-native, plain or flaky, remembering the answer beside the session. A planning model is then given instructions matched to that: fill the schema, or copy this object. The agents page describes how a planning model talks back. - 2.10.74 - 2.10.80 A turn planned out.
/agent <task>answers one turn by planning it first: the chat model splits the task and names the helper each piece goes to, every piece is sent with the conversation's memory selected for it alone, and the chat model writes the reply from what came back. A piece that never answered is shown to it as a gap rather than left out, and a helper's words are quoted on the way in, as material to use rather than instructions to follow. The turn itself is an ordinary turn, kept and remembered like any other.--agent-max-delegationsand--agent-max-wallbound what one round may cost, and every planned turn leaves a line inagent_trace.jsonlthat/statsreads back. - 2.10.82 - 2.10.89 A session that decides its own switches.
--switch-agentwith--switch-rules FILElets a rules file decide the memory switches per turn. A rule is a sentence about the session and the switch it changes while that sentence is true, read by a parser that only compares the numbers a session reports about itself and never runs anything. A file that names a signal nobody reports, a switch a turn cannot set, or two switches known to cancel each other is refused before the session starts. Every decision lasts one turn and is written into nothing, and a turn a rule changed says which rule, what was true and what it set, in/statsand in that turn's own record. A sample ships assalt/agents/switch_rules_sample.jsonwith one rule there is a reason for and two marked as examples that stay unloaded unless--switch-rules-allow-examplesasks for them.--switch-policy modelis the same seam with the chat model proposing instead, experimental, and held to the same refusals. - 2.10.90 - 2.10.101 A reasoning model in the loop. A roster can
name a model for the planning job with
"role": "orchestrator", and a turn then plans with it, under that entry's own settings and under a schema when its server accepts one, with the reply stamped with the model that wrote it. One that is not running costs nothing, since the session's own model plans instead. Pieces bound for different helpers now go out together and come back in the plan's order however they arrived.--agentplans every turn instead of one at a time and marks the reply with a line--agent-quietdrops.--agent-rounds 2lets the orchestrator ask once for one more thing under what is left of the turn's limits. A round where nothing came back answers the turn from the conversation instead of writing up nothing, and one where some pieces failed is told how many before it reads any of them.--offload-ingest-capbounds what a helper's answer may add to memory. Two fixes for reasoning models: a reply whose opening think tag was never in it is cut correctly, and a turn now keeps what a model said rather than what it thought wherever that model sits. - 2.10.102 - 2.10.107 The whole scenario, run end to end and
pinned. Turns files gained
docandagentlines, so a scripted conversation can attach a file and plan a turn the way a person would, andsalt/agents/demo_turns.jsonreplays all four kinds of line deterministically. A session with no roster is pinned to consult nobody, file nothing and cost one call a turn, down to each turn's own record. The architecture pages carry the agent layer. - 2.10.108 - 2.10.119 Hardening across the layer. A parallel round
now stops for its shared token budget the way it stops for its
clock, hands out nothing when a second round has nothing left, and
an interrupt comes back as recorded pieces that keep what arrived
instead of a frozen session. A request's timeout and its usage
numbers travel under the worker's own lock, one delegation to a dead
endpoint no longer condemns the worker, and a server still running
from an earlier session is protected from being spawned over.
Placement now checks the memory a server really takes, refusing a
shared card the declared shares plus resident overhead cannot fit
and reading the card's live free memory when
nvidia-smican be asked. A reply of runaway think tags and JSONNaNorInfinityare refused rather than crashed on, a switch decision cannot combine with the session's own settings into two switches known to cancel, and the tail-occupancy signal counts two messages per exchange. The MCP server takes one call at a time however many arrive, its writes refuse a conversation nobody made,syncwaits the ingest queue out, and every text one call carries meets the same size bound. - 2.10.120 - 2.10.123 The sample rules speak the signals' real
language.
switch_rules_sample.json's examples now key on the verbatim window running light in a long conversation and on orphan mass as the weight of words it is, both able to fire on a real session, and the suite pins that they do.
2.9.0 - 2.9.123
Background ingestion for saltChat. The per-turn keyword and embedding
passes moved off the REPL's critical path onto a worker thread, so long
pasted messages never delay the next prompt. Failed ingests keep the message
text in ingest_failures.jsonl. The session is saved once per turn instead
of once per message. --sync-ingest restores the old inline behavior.
Patch releases, grouped where several versions shipped one thing:
- 2.9.4 - 2.9.9 Persistent serving.
saltServelaunches a long-livedvllm serveprocess,saltChat --backend vllm-serveconnects to it, and a resumed session renders the same prompt bytes (attach order and recent exchanges saved), so it picks up warm. Server errors surface instead of passing off truncated replies. - 2.9.12 - 2.9.16 Multi-GPU.
--gpu 0,1splits a model across cards onsaltServeand both chat backends, the encoder rides the last card, and card order is pinned so an index always means the same physical card. - 2.9.19 Scripted turns.
--turns FILEreplays a JSON or JSONL conversation into one session and--turns-outrecords each answer. - 2.9.21 - 2.9.27 Provenance-aware memory. Every excerpt is labeled
with its turn, speaker and age, and a conversation map (one line per
turn,
--conversation-mapto put it in the prompt) shows what a long session has covered. Refinements keep short sentences from dominating the map and make a partial map say so. - 2.9.31 - 2.9.32 Short turns. Terse decisions like "go with
option B" stay in conversation memory, and
--short-turns fusestores a bare "yes" together with the question it answers. - 2.9.34 - 2.9.36 Verbatim conversation text. Pasted code keeps its
generics, tables and pipelines keep their pipes, link sentences keep
their prose with the URL as
<url>. Older sessions keep their stored text as is. - 2.9.37 - 2.9.38 Interrupted saves. A session whose files disagree
after a crash rolls back to the last complete state on the next open,
with a notice and a record in
load_repairs.jsonl. - 2.9.41 Theme scope.
--per-source-themesprofiles the conversation and each attached file separately, so one large attachment cannot crowd the conversation out of memory. - 2.9.43 - 2.9.47 Memory ceiling.
--memory-cap auto(the default) fits the memory block to the model's window instead of a percentage that grows without bound, and the overflow warning counts the whole prompt. - 2.9.49 - 2.9.53 Stable coverage keys.
/statscounts remembered keys that matched or orphaned each turn, and the opt-in--stable-coverage-keysfreezes the keyword order so remembered discounts survive as the session grows. - 2.9.55 - 2.9.58 Bounded bookkeeping.
/statssplits remembered keys into live and orphaned,--coverage-gccollects the orphans,--coverage-max-keysputs a hard limit on the dictionary, and the code stops claiming a bound the defaults never delivered. - 2.9.76 - 2.9.78 Failure-safe bookkeeping. The per-turn theme discounts, freshness stamps and topic baseline now commit only after the model answers. A turn that errors out leaves memory as it was, so the retry is not fighting a discount from the failed attempt.
- 2.9.82 - 2.9.87 Bounded long sessions.
--max-sentencescaps how many conversation sentences stay in memory, masking the oldest out of selection instead of deleting them, so their text, their numbering and the saved record all survive. Attached files are never masked./statsreports how many sentences are still live. - 2.9.94 - 2.9.96 Incremental compression. Each sentence's lexical tokens and the session's keyword profile are worked out once, when the text arrives, and carried forward instead of being rebuilt from the whole conversation every turn. What memory selects is unchanged, only the work behind it is smaller.
- 2.9.102 Repeating a capped sentence. With
--max-sentenceson, saying something again after the cap masked the old copy away stores it again, instead of dropping it as a copy of a sentence no longer in memory. - 2.9.104 Failure-safe frozen keys. With
--stable-coverage-keyson, a turn whose model call fails leaves the frozen keyword order and its sticky theme set exactly as they were, so a retry does not build on a discarded attempt. - 2.9.106 Corrections in memory. With
--dedup-coson, a fused acknowledgement (--short-turns fuse) is exempt from the gate, so a bare "no" is not dropped as a near-duplicate of the "yes" it corrects. A short reversal built from the same words on the default path stays a documented limit of the cosine gate. - 2.9.108 - 2.9.111 Tail-aware memory selection. The memory block
no longer spends its budget on sentences the model is already reading
verbatim in the recent messages, so the budget buys older context
instead of repeating what is on screen. A sentence's themes start
counting as shown once it leaves the recent messages. On by default,
--no-tail-excluderestores the old overlapping selection, and/statsreports how many sentences were left out.
2.8.0
Near-duplicate memory gate. Opt-in --dedup-cos skips a new conversation
sentence too similar to an earlier one from the same speaker, so
restatements stop inflating theme statistics. Each suppression is logged to
near_dups.jsonl and counted in /stats.
2.7.0
PDF tables reach the model: table and pseudocode rows are grouped under their captions as caption-prefixed units, so numbers stay readable against their column names. Sturdier PDF extraction overall.
2.6.0 - 2.6.7
vLLM chat backend. saltChat --backend vllm serves registered weights
through an in-process vLLM engine with automatic prefix caching, streaming
tokens through the same REPL. Real prefix-cache hits are recorded in the
kvtrace ledger next to the selection split. --gpu-mem-util and
--max-model-len control the engine. Includes a regression harness.
2.5.0
Topic-shift damping. Opt-in --shift-damping detects when a question pivots
away from the recent conversation and scales down stale theme suppression
for that turn only, with --shift-margin and --shift-query-boost as
tuning knobs. Drift readings always show in /stats.
2.4.0
Coverage decay. Opt-in --coverage-half-life lets a surfaced theme's
suppression fade over turns of silence, so topics the conversation returns
to can resurface. --coverage-decay-docs opts attached files in.
2.3.0 - 2.3.1
kvtrace: a per-conversation KV read/write ledger (events.jsonl plus a
per-token matrix) recording reused, fresh, and output tokens every turn.
- 2.3.1 PDF-to-sentence pipeline (headers, ligatures, reflow, reference filtering), source-labeled memory blocks, and the KV-cache-shaped prompt layout with block-wise tail compaction.
2.2.0 - 2.2.1
salt@ attachments: staged files ingest into their own trie branch so
multiple attachments never crowd each other out.
- 2.2.1
attach@full-context attachments: a file's whole text rides uncompressed in every prompt.
2.1.0
Chatbot mode. The saltChat REPL with a persistent per-conversation trie,
plus the salt console command.
2.0.0
CELF coverage selection became the default selector.
1.0.0
First release: LongBench compression and evaluation.