Use it in 30 seconds
Each MCP setup below starts the same server, llmslim mcp, with uv (Python 3.10 or
newer). Memory goes to a local SQLite file, ~/.llmslim/memory.db.
One click:

| Tool | Run |
|---|
| Claude Code | claude mcp add llmslim -- uvx --from "llmslim[mcp]" llmslim mcp |
| Claude Code plugin | /plugin marketplace add Thanatos9404/llmslim, then /plugin install llmslim@llmslim |
| Codex | codex mcp add llmslim -- uvx --from "llmslim[mcp]" llmslim mcp |
| GitHub Copilot CLI | copilot mcp add llmslim -- uvx --from "llmslim[mcp]" llmslim mcp |
| Gemini CLI | gemini extensions install https://github.com/Thanatos9404/llmslim |
| Hermes Agent | hermes plugins install llmslim --enable, then hermes config set context.engine llmslim: a context engine that compacts long conversations locally (plugin) |
The first start downloads LLMSlim and its dependencies. If a client gives up waiting, run
uvx --from "llmslim[mcp]" llmslim mcp --help once to fill uv's cache, then restart the client.
Python (pip install llmslim, Python 3.10 or newer):
from llmslim.memory import Memory
mem = Memory("memory.db")
mem.add([{"role": "user", "content": "I'm allergic to peanuts."}], user_id="asha")
context = mem.search("Am I allergic to anything?", user_id="asha", budget=1500)
print(context.text)
TypeScript (npm install llmslim, Node 18 or newer), against llmslim serve (pip install "llmslim[server]"):
import { LLMSlim } from "llmslim";
const slim = new LLMSlim();
await slim.memories.add({ userId: "asha", messages: "I'm allergic to peanuts." });
const context = await slim.memories.search({ userId: "asha", query: "Am I allergic to anything?", budget: 1500 });
Runnable examples, each with a test: Claude Code memory,
LangChain agent, Pydantic AI agent,
big logs and TypeScript.
If LLMSlim is useful, a star helps others find it. Feedback is welcome in issues.
What it does
- Five MCP tools:
remember, recall, list_memories, forget and fit_context.
- Stores messages word for word in SQLite, Postgres or MongoDB. Nothing is summarized or extracted on the way in.
- Recalls within a token budget.
recall returns the stored messages that answer a question, under their dates
and whole where they fit, within the budget you set. fit_context shrinks a long log or document to a budget.
- Fixes the scope at server start. The user and agent ids are set when the server starts, not by the model, so a
model steered by injected text can't read or delete another user's memory.
- Makes no LLM call. By default, storing and recalling are plain code: no model, no embedding API, no API key.
- MIT licensed. The library and the MCP server need Python 3.10 or newer.
Beyond memory, LLMSlim Core also compresses and plans prompts, chat history, retrieved documents and tool schemas
into token budgets: see the sections below.
Why no LLM call
Storing is a database write and recall is deterministic ranking code. So what comes back is what was said, not a
model's rewrite of it, and the same history and question always give the same context, which makes recall easy to
test. The memory server and llmslim serve make no network calls of their own. What does reach the network:
installing from PyPI (including uvx's first run of a version) and tiktoken downloading its tokenizer file once.
The trade-off: the default lexical mode matches words, not meanings, so a question phrased differently from the
stored message can miss it. relevance="semantic" adds an optional cross-encoder for that (llmslim[semantic],
which downloads the model the first time). See the limits.
Compress a prompt in 60 seconds
1. Install LLMSlim Core from PyPI (Python 3.10+). Running it needs no API key, model, or network call:
2. Run this snippet:
from llmslim import compress
ticket = """
Hi team, I hope you are all doing well and had a good weekend.
I am writing about invoice INV-2291, which charged us $480 instead of $240.
We upgraded to the Pro plan on 3 September, but the old Basic plan was billed as well.
Our finance lead was asking about this during the weekly sync, which ran long as usual.
Please refund the duplicate $240 charge to the card ending 4417.
Thanks so much, and let me know if you need anything else from me.
"""
result = compress(ticket, target_ratio=0.6)
print(result.compressed_text)
print(f"{result.original_tokens} -> {result.compressed_tokens} tokens ({result.reduction_percent:.1f}% fewer)")
3. What you should see (llmslim 0.11.1, base install):
Hi team, I hope you are all doing well and had a good weekend. I am writing about invoice INV-2291, which charged us $480 instead of $240. We upgraded to the Pro plan on 3 September, but the old Basic plan was billed as well. Please refund the duplicate $240 charge to the card ending 4417.
112 -> 72 tokens (35.7% fewer)
Two of the six sentences are dropped; the invoice number, amounts, plan change,
and refund request are kept. The default extractive mode only selects sentences
from your input, so the same input always gives the same output. The base install
estimates tokens at about 4 characters per token and prints a one-line warning
saying so. With pip install "llmslim[fast-tokens]", tokens are counted with
tiktoken instead: on this input the same four sentences are kept, and the count
reads 107 -> 72 tokens (32.7% fewer).
Next: Docs • Pricing • Vault beta • Issues
LLMSlim Core, Orbit and Vault
The context layer for production AI agents.
This repository is LLMSlim Core (v0.12.0, MIT): local-first context planning and compression that fits prompts, conversation history, RAG, memory, and tool schemas into real token budgets while preserving instructions and execution boundaries.

| LLMSlim Core | LLMSlim Orbit | LLMSlim Vault |
|---|
| What it is | The open-source library, MCP server and HTTP server | The hosted edition, run by us | The self-hosted edition, run in your own VPC (formerly LLMSlim Platform) |
| Version | 0.12.0 | Opens with LLMSlim 1.0 | 0.9.1 beta |
| License | MIT, this repository | Proprietary, paid plans | Proprietary, private beta |
| Get it | pip install llmslim | Join the founding list at llmslim.app/pricing | By invitation; not published to PyPI |
| What it does | Compresses, plans, and traces model-visible context inside your Python process | A hosted memory API, remote MCP with OAuth for Claude.ai, ChatGPT, Cursor and Claude Code, team memory, a context console with traces and replay, and a truth timeline | Builds on Core: a verifiable timeline of agent state, explainable and replayable context decisions, and outcome-driven context policies that start in shadow mode |
Core is and stays fully open source. Orbit and Vault are separate products that depend on
Core; none of their code lives in this repository. Plans and prices are at
llmslim.app/pricing; read about Vault and request beta access at
llmslim.app/vault.
What's new in Core: the agent context runtime (v0.7.0)
and the cache-aware context runtime (v0.7.1).
Both are in the PyPI package from 0.7.1 onward; 0.7.0 was tagged but not published to PyPI.
Adaptive Context Planner (v0.6.0)
LLMSlim is no longer only a fixed-ratio prompt compressor. The Adaptive
Context Planner chooses safe representations for each context item, then uses
a deterministic constrained allocator to fit the highest-value context into a
finite model budget. The original compress() API remains fully supported.
The v0.6.0 release adds explainable per-item decisions, explicit infeasibility
reporting, post-plan validation, token and cost telemetry, policy presets, and
optional Sarvam AI, Zoho, and MongoDB integrations while keeping the base
installation local-first and provider-neutral.
from llmslim import plan_context
plan = plan_context(
messages=[
{"role": "system", "content": "Answer only from verified context."},
{"role": "user", "content": "When does Acme renew?"},
],
documents=[{"content": "CRM: Acme renews on 2026-11-30."}],
query="Acme renewal date",
model="sarvam-105b",
max_input_tokens=8_000,
reserve_output_tokens=512,
)
assert plan.feasible
print(plan.final_context)
print(plan.metrics.planned_tokens, plan.metrics.estimated_input_cost_after)
Planning works offline with the base install. Optional integrations are
isolated behind extras:
pip install "llmslim[sarvam]"
pip install "llmslim[zoho]"
pip install "llmslim[mongodb]"
See the architecture,
algorithm, security model,
CLI, and measured offline benchmark.
Startup programs
LLMSlim is part of 21 startup programs. They provide credits, tools, and guidance;
none of them is an investor, and membership does not imply endorsement.
Logos are trademarks of their owners and identify each program.
Sarvam Startup Program and Sarvam integration
✦ Accepted into the Sarvam Startup Program
LLMSlim has been accepted into the Sarvam Startup Program. This is program
membership only: it is not a partnership, and it does not imply an endorsement
by Sarvam AI.
Separately, LLMSlim Core ships an optional Sarvam integration (llmslim[sarvam])
built on the official sarvamai Python SDK. You plan and compress context locally,
then send only the planned context to a Sarvam model such as sarvam-105b.
- Local planning: retrieved documents are planned and compressed in your Python process before any request is sent.
- Local-first: sensitive retrieved context stays local; only the planned context is sent to the model.
- Your application stays in charge: LLMSlim is a pre-processor. Your application owns API keys, prompts, and inference, and nothing is sent to Sarvam unless your code calls the provider.
Quickstart: Adaptive Planning + Sarvam
pip install "llmslim[sarvam]"
export SARVAM_API_KEY="..."
from llmslim import plan_context
from llmslim.integrations.sarvam import SarvamProvider
documents = [
{"content": "Q3 filing: new data-localisation rules take effect in April 2027."},
{"content": "Q3 filing: two pending licence renewals carry a regulatory risk of delay."},
]
plan = plan_context(
messages=[{"role": "system", "content": "Answer from verified context."}],
documents=documents,
query="Summarize the key regulatory risks.",
model="sarvam-105b",
max_input_tokens=8_000,
reserve_output_tokens=256,
)
provider = SarvamProvider.from_env(model="sarvam-105b", max_tokens=256)
answer = provider.chat([{"role": "user", "content": plan.final_context}])
print(answer, provider.last_usage)
👉 Read the Sarvam integration guide (also in docs/integrations/SARVAM.md)
🔮 What is LLMSlim?
LLMSlim Holographic Engine: Multi-stage token compression condensing sparse context streams into dense information cores with zero contract mutation.
Modern LLM systems suffer from the Context Window Dilemma:
- Financial Tax: Re-sending large prompts on every turn means you pay for the same input tokens again and again.
- Latency Bottleneck: Large prompts inflate Time-To-First-Token (TTFT) and queue times.
- Lost-In-The-Middle Syndrome: Models can miss or misuse information buried in the middle of long contexts.
- Security Risks: Untrusted RAG content can smuggle indirect prompt injections into system instructions.
LLMSlim fixes this at the source. It provides deterministic, contract-preserving compression that runs natively in your application layer before calling any model.
💎 Core Architectural Pillars
| Pillar | How LLMSlim Solves It |
|---|
| 🏎️ Ultra-Low Overhead | Pure Python + NumPy/Scikit-learn. Median 0.026 ms per sample (p95 1.889 ms) in the 24-sample, mostly short-text Phase 2 evaluation, with no model calls on the default path. Longer inputs take longer. |
| 🔒 Deterministic by Default | The default extractive mode only selects sentences from your input, so output is reproducible and contains no generated text. It can still drop a fact you need, so check results for your use case. Rewrite and hybrid modes call your own LLM. |
| 🛡️ Provenance Security | ContextRole hierarchy isolates untrusted RAG/tool text so prompt injections cannot escalate to system priority. |
| 🧩 Contract-Safe Tool Schemas | Canonical JSON normalization, SHA-256 contract fingerprints, and zero execution rewrites for agent tooling. |
| 🔌 Provider-Neutral | Execution-free request builders for Sarvam, OpenAI, Anthropic, Gemini, and vLLM, plus an OpenAI Agents SDK bridge. Works with any framework that can pass plain message or document lists. |
Paste a long chat, a stack of search results or a CI log, ask a question, and watch LLMSlim keep what
answers it. The Studio shows the tokens and cost saved, whether the answer survived, and what plain truncation to
the same size would have kept instead.
- Ready-made scenarios: a six-month support chat, help-centre search results, and a CI log with one failure,
each with a question whose answer is known.
- Your own text: paste up to about 16,000 tokens and choose how much to keep.
- In place: kept lines are highlighted inside the original, next to the exact text the model would receive.
- No model call: the Studio runs Core on the server and calls no provider. Nothing you paste is stored.
See the security policy. The website's source, the Studio included, lives in a private repository;
Core, everything pip install llmslim installs, is here under the MIT license.
👉 Open Interactive Studio
🌐 Modern Developer Website & Documentation
Explore our redesigned dark/light documentation site covering end-to-end integration patterns:
📦 Installation
pip install llmslim
pip install "llmslim[mcp]"
pip install "llmslim[agents]"
pip install "llmslim[semantic]"
pip install "llmslim[sarvam]"
pip install "llmslim[exa]"
pip install "llmslim[neo4j]"
pip install "llmslim[zoho]"
pip install "llmslim[mongodb]"
pip install "llmslim[langchain]"
pip install "llmslim[llamaindex]"
pip install "llmslim[haystack]"
pip install "llmslim[gptme]"
pip install "llmslim[pydantic-ai]"
pip install "llmslim[all]"
Requirements: Python 3.10 or higher. No provider SDK, PyTorch, sentence-transformers, or cloud service is required for the default installation.
⚡ Quick Start & Usage Patterns
from llmslim import ContextRole, compress
text = """
The Apollo program was conceived in 1960 during the Eisenhower administration.
NASA announced the program as a follow-up to Project Mercury.
The spacecraft was designed to carry three astronauts.
Apollo succeeded in landing the first humans on the Moon in 1969.
Funding for the program peaked in 1966 with over $4.5 billion appropriated.
"""
result = compress(
text,
target_ratio=0.5,
strategy="extractive",
context_role=ContextRole.GENERAL,
)
print("Compressed Output:")
print(result.compressed_text)
print(f"Original: {result.original_tokens} tokens | Compressed: {result.compressed_tokens} tokens")
print(f"Saved: {result.tokens_saved} tokens ({result.reduction_percent:.1f}%)")
print(f"Sentences kept: {result.sentences_kept} of {result.sentences_total}")
Extractive selection keeps whole sentences, not meaning: always check that the
sentences you need survived. On this text, the 0.7.1 base install keeps three of
the five sentences and drops the 1969 Moon landing sentence; with
llmslim[fast-tokens] (tiktoken counts) it keeps that sentence and drops another.
2. Multi-Document RAG Compression
Compress each retrieved chunk before you build the prompt. compress_documents
takes plain strings and returns one result per document, in order:
from llmslim import compress_documents
documents = [
"Q3 revenue grew 23% year-over-year to $1.2B. "
"Growth came mainly from enterprise subscriptions in India and Southeast Asia. "
"The board met twice during the quarter. "
"The office cafeteria menu was updated in August. "
"A new logo was unveiled at the annual offsite. "
"Net revenue retention reached 118%, up from 109% a year earlier.",
"Operating expenses rose 11% because of AI research infrastructure. "
"Most of the increase was GPU capacity for model evaluation. "
"Travel costs were flat. "
"The parking garage was repainted in September. "
"Employees voted on a new name for the meeting rooms. "
"Headcount in engineering grew by 40 people.",
]
results = compress_documents(
documents,
query="What drove Q3 revenue and cost changes?",
target_ratio=0.5,
)
context = "\n\n".join(r.compressed_text for r in results)
print(context)
print(f"Saved {sum(r.tokens_saved for r in results)} tokens across {len(results)} documents")
With a query, the documents share one budget and it goes where the query points. Relevance is lexical by default
(BM25 with feedback: no model, deterministic, tens of milliseconds on a 3k-token context). For RAG and long chat
memory, where the answer often shares few words with the question, add the optional reranker:
results = compress_documents(documents, query=question, target_ratio=0.25, relevance="semantic")
relevance also accepts any callable scorer(query, texts) -> scores, such as a hosted reranking API.
Retrieving with LangChain, LlamaIndex or Haystack? Use the
framework adapters instead of calling
compress_documents yourself.
NOTE
compress_documents treats documents as untrusted (ContextRole.RAG) by default,
so imperative wording inside a retrieved document is never force-kept as a
protected instruction. It does not remove injected text: an injected sentence
can still survive compression. To mark retrieved content as untrusted in the final
prompt, use plan_context or
ContextRuntime, which wrap every document in a
trusted="false" provenance tag and ignore any role a document claims for itself.
3. Multi-Turn Chat Conversation Compression
Compress long earlier turns while passing system instructions and user questions through unchanged:
from llmslim import compress_chat_messages
review = (
"Here are my suggestions for your schema. First, add a composite index on "
"(customer_id, created_at) because most of your queries filter by customer and sort by "
"date. Second, the orders table stores the shipping address inline, which duplicates "
"data across rows; move it to an addresses table. Third, the status column is a "
"free-text string, so use an enum or a lookup table instead. Fourth, the price column "
"uses FLOAT, which loses precision for money; use DECIMAL(12, 2). Finally, add ON "
"DELETE RESTRICT to the foreign key from orders to customers so that deleting a "
"customer cannot silently orphan their orders."
)
messages = [
{"role": "system", "content": "You are a specialized code review engineer."},
{"role": "user", "content": "Can you review my database schema design?"},
{"role": "assistant", "content": review},
{"role": "user", "content": "What about the foreign key constraint on line 42?"},
]
compressed_messages = compress_chat_messages(
messages,
target_ratio=0.5,
compressible_roles=("assistant",),
min_tokens=60,
)
for message in compressed_messages:
print(f"{message['role']}: {message['content']}")
Combine ultra-fast extractive pre-filtering with your own LLM provider for deep semantic paraphrasing:
from llmslim import CallableProvider, compress
import openai
client = openai.OpenAI()
long_raw_text = """
The Apollo program was conceived in 1960 during the Eisenhower administration.
NASA announced the program as a follow-up to Project Mercury.
The spacecraft was designed to carry three astronauts.
Apollo succeeded in landing the first humans on the Moon in 1969.
Funding for the program peaked in 1966 with over $4.5 billion appropriated.
"""
provider = CallableProvider(
lambda req: client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": req.system_prompt},
{"role": "user", "content": req.user_prompt},
],
).choices[0].message.content
)
result = compress(
long_raw_text,
target_ratio=0.3,
strategy="hybrid",
provider=provider,
)
meta = result.rewrite_metadata
print(result.compressed_text)
print(f"Strategy: {meta.strategy} | Rewrite accepted: {meta.accepted} | Fell back to extractive: {meta.fallback_used}")
print("Validation failures:", meta.failure_reasons)
Every rewrite is validated for structure, instruction retention, entity retention,
and similarity. If a hybrid rewrite fails validation, LLMSlim returns the extractive
result instead and reports why in rewrite_metadata.failure_reasons.
Coding agents: llmslim wrap claude
Claude Code resends the whole conversation on every turn. So long test logs, build output, search results and web
pages get paid for again and again, and they add to the context that Claude Code eventually compacts.
llmslim wrap claude runs Claude Code through a local proxy:
- It shrinks such output the first time the model sees it. Lines that look like errors, failures or warnings come
first, then the lines related to the command.
- It sends exactly the same bytes on every later request, so the prompt cache keeps hitting and Claude's thinking
stays valid.
- Output of Claude Code's own
Read, Edit and Write tools is never touched. A file printed through the shell
(cat, head) counts as shell output.
llmslim wrap and llmslim proxy are experimental: they may change in a minor release
(API stability).
pip install "llmslim[proxy]"
llmslim wrap claude
llmslim proxy stats
What it does to the bill, measured on real Claude Code 2.1.296 with Claude Haiku 5.5, on eight small synthetic
tasks. We counted list-price dollars from the API's usage, with cache reads priced separately, and published all 572
runs in benchmarks/claude_code_cost. The primary result has both arms in
Claude Code's default permission mode; a second run has both in auto mode.
- Long logs: a simulated failing CI run (about 12 KB of log, the cause in the middle) cost 26-32% less, and a lint
run with three errors among 240 warnings 17-21% less.
- Test runs: about 4% less in default mode, within noise for every task; 2-15% less in auto mode. Agents already
trim test output with
tail.
- Over 8 tasks: +5.3% in default mode (95% interval +1.5% to +9.4%), mostly from one task where shortened JSON
cost about twice as much; -5.1% in auto mode. JSON is left alone by default since 0.11.1.
- Real sessions: in an offline replay of the maintainer's own transcripts (32 sessions and the 412 subagent runs
they started, almost all Claude Opus 5.5, at list prices), tool output was a small part of the bill: with the same
turns, all shell, search and web output costing nothing would have cut it by 13.4%. Re-reading long context was the
big part. Claude Code's
/autocompact setting decides when it compacts; what that does to cost is not measured
here.
Without the byte-for-byte replay, an edited history invalidates Claude's thinking: in our API check on Claude Haiku
5.5 the request was rejected with HTTP 400, the default for accounts created on or after 31 August 2026. With the
replay, cache reads grew on every request of a 7-request API test. See docs/PROXY.md for what is compressed, the
limits and security.
As a Claude Code plugin. The plugin adds LLMSlim's memory server (below) and two commands:
/llmslim:wrap-claude walks you through starting Claude Code through the proxy;
/llmslim:savings shows how many tool-output tokens the proxy kept out of your requests: an estimate counted
before prompt-cache discounts, not dollars saved.
/plugin marketplace add Thanatos9404/llmslim
/plugin install llmslim@llmslim
On Claude Code 2.1.275 or later, /plugin install llmslim --marketplace Thanatos9404/llmslim does both in one step.
The memory server runs with uvx, so it needs uv.
Memory for agents, with no LLM in the loop
llmslim.memory stores every turn verbatim and selects, at read time, the turns that answer a question within a
token budget. There is no extraction model on the way in and no model call on the way out.
from llmslim.memory import Memory
memory = Memory("memory.db")
memory.add([{"role": "user", "content": "I'm allergic to peanuts."}], user_id="asha")
context = memory.search("Am I allergic to anything?", user_id="asha", budget=1500)
print(context.text)
On BEAM's 1M-token conversations (held-out, run once), Memory.search with its reranker scored on par with Mem0 at
2,500 and 10,000 tokens (+1.2 and +0.6 points, neither significant), 60x and 28.6x cheaper per 1,000 questions at
list prices, and 7.5 to 9.9 points ahead of Zep (on 30 conversations). Its weak spots are published alongside the results
(Round 6). Coming from Mem0?
from llmslim.memory.compat import Mem0Memory as Memory keeps Mem0's method names, and llmslim import mem0 export.json brings your memories along. llmslim import also reads Claude Code transcripts and ChatGPT and
Claude.ai exports (docs/IMPORT.md). See
docs/MEMORY.md for scopes, stores, evidence and limits, and the 1.0 roadmap.
In Claude, Cursor or any MCP client
llmslim mcp serves this memory over MCP, with a fit_context tool for long text. It needs pip install "llmslim[mcp]"
on Python 3.10+.
claude mcp add llmslim -- llmslim mcp
- Tools:
remember, recall, list_memories, forget and fit_context.
- Scope: the server fixes it with
--user-id, so the model can't read another user's memory.
- Storage: a local SQLite file; the server makes no network calls of its own (the tokenizer downloads one file once and caches it).
- Several machines (experimental):
llmslim mcp --transport http (1.0 and newer) serves the same memory over
HTTP behind bearer keys, so your laptop and desktop share it; see
docs/MCP.md.
- More: client configs and options are in docs/MCP.md.
One-click install in other clients. Each link runs uvx --from "llmslim[mcp]==0.12.0" llmslim mcp, so it needs
uv and Python 3.10 or newer.

In Gemini CLI, gemini extensions install https://github.com/Thanatos9404/llmslim adds the same server.
In VS Code, the LLMSlim Memory extension adds it to VS Code's MCP support (coming soon to the VS Code
Marketplace and Open VSX).

Over HTTP, from any language
llmslim serve puts the same memory and a context-fitting endpoint behind a small JSON API. It runs as one process with
one SQLite file and no model calls. Install it with pip install "llmslim[server]", or build the Dockerfile.
curl -s localhost:8080/v1/memories/search -H 'content-type: application/json' -d '{"user_id": "asha", "query": "Where did I move to?", "budget": 1500}'
It needs a bearer API key to listen beyond localhost. The routes, limits and Docker setup are in
docs/SERVER.md.
The TypeScript client, npm install llmslim (clients/typescript), wraps this API with no
dependencies. It also gives the Vercel AI SDK recall_memory and remember tools and a middleware that saves turns. None of them changes your
prompt, so prompt caches keep hitting and Claude's preserved thinking stays valid.
Use with LangChain / LlamaIndex / Haystack
Each adapter runs compress_documents on the retrieved documents, with the same
shared budget when there is a query. Kept documents keep their metadata, and
documents that compress to nothing are dropped. They run locally with no model.
| Framework | Install | Class | Use it where you'd use |
|---|
| LangChain | llmslim[langchain] | llmslim.integrations.langchain.LLMSlimCompressor | LLMLinguaCompressor, e.g. in ContextualCompressionRetriever |
| LlamaIndex | llmslim[llamaindex] | llmslim.integrations.llama_index.LLMSlimPostprocessor | LongLLMLinguaPostprocessor, in node_postprocessors |
| Haystack | llmslim[haystack] | llmslim.integrations.haystack.LLMSlimDocumentCompressor | a component between the retriever and the prompt builder |
from langchain_classic.retrievers.contextual_compression import ContextualCompressionRetriever
from llmslim.integrations.langchain import LLMSlimCompressor
retriever = ContextualCompressionRetriever(
base_compressor=LLMSlimCompressor(target_ratio=0.25),
base_retriever=your_retriever,
)
Runnable examples with their output:
LangChain,
LlamaIndex,
Haystack.
Use with gptme / Hermes Agent / Pydantic AI
These adapters fit an agent's message history into a token budget. They keep the
system prompt, tool calls and the latest turns as they are, compress older tool
output and turns locally (failures and tracebacks in tool output are kept), and keep
every tool call paired with its result. There is no model call.
| Agent | Install | Class or function | How it plugs in |
|---|
| gptme | llmslim[gptme] (gptme 0.34+) | llmslim.integrations.gptme.LLMSlimContextProvider | gptme context provider llmslim; [plugin.llmslim] compaction = true makes it the one gptme's compaction uses |
| Hermes Agent | llmslim in Hermes's environment | llmslim.integrations.hermes.LLMSlimContextEngine | ContextEngine plugin llmslim; select it with context.engine: llmslim |
| Pydantic AI | llmslim[pydantic-ai] (1.0+) | llmslim.integrations.pydantic_ai.llmslim_history_processor | history processor, in ProcessHistory(...) or history_processors=[...] |
from pydantic_ai import Agent
from pydantic_ai.capabilities import ProcessHistory
from llmslim.integrations.pydantic_ai import llmslim_history_processor
agent = Agent(
"openai:gpt-5",
capabilities=[ProcessHistory(llmslim_history_processor(max_tokens=8_000))],
)
Setup and runnable examples with their output:
gptme,
Hermes Agent,
Pydantic AI.
🛡️ Security Architecture: Provenance Isolation
Compression algorithms that naively score sentences based on imperative keywords ("Must", "Always", "Action Required") introduce a severe vulnerability: indirect prompt injection amplification. An attacker injecting malicious instructions into a retrieved RAG document can trick the compressor into preserving the attack payload while discarding legitimate context.
We found this issue in LLMSlim internally and fixed it in v0.3.1. No CVE or
security advisory was published, and the issue has no official CVSS score.
LLMSlim solves this via an explicit Context Role Trust Boundary:
┌────────────────────────────────────────────────────────┐
│ HIGHEST TRUST (Priority Tier 4) │
│ System & Developer Prompts (Strictly Protected) │
├────────────────────────────────────────────────────────┤
│ INTERMEDIATE TRUST (Priority Tier 2-3) │
│ Direct User Input │
├────────────────────────────────────────────────────────┤
│ PROVENANCE-LOCKED (Priority Tier 0-1) │
│ RAG Documents • Assistant History • Tool Outputs │
│ *Cannot reach Tier 4 through imperative patterns* │
└────────────────────────────────────────────────────────┘
- Hardened Priority Locking: Content marked
ContextRole.RAG, ContextRole.TOOL, or ContextRole.ASSISTANT cannot achieve Tier-4 priority, regardless of imperative syntax.
- Nonce-Protected Rewrite Delimiters: When the input contains a fence-like token such as
---END TEXT---, the rewrite template's fences get a random per-call nonce, so the embedded token cannot close the text region. The input itself is never modified.
- Untrusted by default, not filtered: these protections stop injected text from gaining priority. They do not delete it.
plan_context and ContextRuntime label it trusted="false" in the final context so your model and application can treat it accordingly.
Large agent catalogs (dozens or hundreds of tool definitions) waste thousands of tokens before an agent even starts thinking. LLMSlim introduces Contract-Safe Tool Infrastructure (v0.4.0 & v0.5.0):
from llmslim.tools import (
canonical_json,
contract_equivalent,
fingerprint_tool_schema,
from_mcp_tool,
optimize_tool_schema,
)
raw_tool = {
"name": "enterprise.search_ledger",
"description": "Searches financial transaction records by date range.",
"inputSchema": {
"type": "object",
"properties": {"account_id": {"type": "string"}, "limit": {"type": "integer"}},
"required": ["account_id"],
},
}
tool = from_mcp_tool(raw_tool, namespace="finance")
result = optimize_tool_schema(tool)
assert result.equivalence.status.value == "EXACT"
assert fingerprint_tool_schema(tool) == fingerprint_tool_schema(result.optimized)
print("Canonical Fingerprint:", fingerprint_tool_schema(result.optimized))
print(canonical_json(result.optimized.raw))
Async MCP Catalog Source (v0.5.0)
Requires pip install "llmslim[mcp]" (Python 3.10+) and an MCP server you run:
import asyncio
from llmslim.mcp import MCPToolCatalogSource, PlanMode, plan_catalog_context
async def main() -> None:
source = MCPToolCatalogSource.from_streamable_http(
"http://127.0.0.1:8000/mcp",
headers={"Authorization": "Bearer local-token"},
)
snapshot = await source.list_tools()
plan = plan_catalog_context(snapshot, mode=PlanMode.FULL)
print(f"Catalog contains {len(snapshot.tools)} tools ({plan.metrics.catalog_tokens} tokens)")
asyncio.run(main())
📊 Benchmarks & Radical Transparency
Many libraries make unsubstantiated claims about 90% prompt compression. At LLMSlim, we practice Radical Benchmark Transparency:
Round 6: agent memory on BEAM-1M, Core 0.12 against Mem0 and Zep
Measured 24-27 October 2026 on held-out data: BEAM's 1M tier, 35 conversations of about 946,000 tokens each and
700 questions across 10 types. None of it was opened while 0.12 was tuned. The protocol was fixed before the first
paid call, its one change to the ship rule and the claim wording was made before any system had answered a question,
and every system was read and judged once.
Every system is read by Claude Haiku 5.5 with BEAM's answer prompt, three times per question, and scored by BEAM's
rubric judge (also Haiku 5.5). The score is the mean rubric score out of 100. Mem0 (OSS 2.2.1, extraction on) and
Zep (Graphiti 0.30.2) stored every exchange through their own pipelines, with Haiku 5.5 as their LLM and
Qwen3-Embedding-0.6B as their embedder. LLMSlim ran with its reranker (relevance="semantic"). Cost is per 1,000
questions at list prices, with BEAM's 20 questions per conversation. It counts the answer call, Mem0's and Zep's
memory-building calls, and each system's compute.
| System | Context tokens (mean) | Score out of 100 (95% CI) | Cost per 1,000 questions |
|---|
| LLMSlim 0.12 + reranker, 2,500 budget | 2,436 | 44.6 (41.9 to 47.3) | $1.22 |
| Mem0, top 50 memories | 3,020 | 43.4 (40.9 to 45.8) | $73.12 |
| LLMSlim 0.12 + reranker, 10,000 budget | 9,766 | 47.3 (44.3 to 50.3) | $2.61 |
| Mem0, top 200 memories | 12,095 | 46.7 (44.1 to 49.3) | $74.53 |
| LLMSlim 0.11.1 + reranker, 2,500 / 10,000 | 2,493 / 9,985 | 38.0 / 41.6 | $0.91 / $2.21 |
| Zep, on 30 of the 35 conversations* | 3,397 | 38.4 (35.6 to 41.2) | $330.07 |
* Zep ran on a seeded draw of 30 conversations to cap its cost. On those 30, LLMSlim 0.12 scored 45.9 and 48.3.
What it shows:
- On par with Mem0, at a fraction of the cost. LLMSlim is +1.2 points at 2,500 tokens (95% CI -0.7 to +3.2) and
+0.6 at 10,000 (-1.6 to +2.7). Neither difference is significant, so we do not claim to beat Mem0. It is 60x and
28.6x cheaper per 1,000 questions at list prices (2,500 and 10,000 tokens), because Mem0 makes an LLM call for
every exchange it stores and LLMSlim makes none.
- Ahead of Zep by 7.5 to 9.9 points, on Zep's 30 conversations: +7.5 at 2,500 tokens (+4.5 to +10.4) and +9.9 at
10,000 (+6.8 to +12.9).
- Ahead of 0.11.1. On the same questions 0.12 scored 44.6 and 47.3 where 0.11.1 scored 38.0 and 41.6, with no
question type more than 1.3 points lower.
Where LLMSlim loses (against Mem0, per question type, 70 questions each):
- Summarization: -8.9 points at 2,500 tokens (-13.1 to -4.7) and -4.0 at 10,000 (-7.8 to -0.2).
- Multi-session reasoning at 10,000 tokens: -6.0 (-10.1 to -1.8).
- Information extraction: -4.6 and -4.9. Neither is significant.
- Search latency with the reranker. With the reranker running on a laptop CPU, a search over a 946,000-token
history took 3.6 s at the median at 2,500 tokens and 4.9 s at 10,000. The default lexical search, which
llmslim mcp uses, takes about 45 ms on a 100,000-token history, but it was not scored in this round.
- The ship rule was relaxed. It was loosened before any LLMSlim result on this data was read. The original,
stricter per-type bar, which also required the lower bounds for instruction following and summarization to stay
within 5 points of Mem0, would have failed on summarization (both budgets) and instruction following (10,000
tokens). The report gives both verdicts.
"LLMSlim 0.12" is commit 16a2708, which became 0.12.0 after hardening and returns the same bytes on the English parity
checks; the round-6 contexts were not rebuilt with 0.12.0. Setup, every table, paired differences, cost, latency and
limitations: benchmarks/live/round6. The protocol, harness and per-question
answers are not published with this round.
Round 3: Core 0.9.0 against memory systems and compressors, cost first
Measured 13-15 October 2026 on held-out test questions. Every system is read by Claude Haiku 5.5 with the same
answer prompt and judges. Mem0 and Zep ran through their own benchmark harnesses, and LLMLingua-2 and LongLLMLingua
(LLaMA-2-7B in float16) with their published settings. Every compressor's context was built in public Kaggle
notebooks pinned to a commit. Each cell is accuracy and what 1,000 questions cost end to end at list prices: the
answer call, Mem0's and Zep's memory-building calls, and each system's compute. Full tables, confidence intervals,
weak spots and every answer: benchmarks/live/round3.
| Benchmark | LLMSlim 0.9 + reranker | LLMSlim 0.9 (no model) | Mem0 | Zep | LongLLMLingua | LLMLingua-2 | Retrieval baseline* | Full context |
|---|
| LoCoMo, 1,307 q, ~2,500 tokens | 93.0% · $0.95 | 91.7% · $0.89 | 91.0% · $2.66 | 74.4% · $6.24 | 73.3% · $7.01 (300 q) | 34.1% · $0.43 | 85.2% · $0.90 | 95.1% · $0.83 cached |
| LongMemEval-S, 180 q, ~10,000 tokens | 74.4% · $2.21 | 77.8% · $2.13 | | | not run | 15.6% · $8.43 | 80.6% · $4.45 | 77.8% · $90.21 |
| LongMemEval-S, Mem0's 60 q, ~2,500 tokens | 78.3% · $1.04 | 71.7% · $0.96 | 76.7% · $220.88 | | | | 80.0% · $3.36 | 78.3% · $90.20 |
| LongMemEval-S, Zep's 24 q, ~2,500 tokens | 87.5% · $1.01 | 87.5% · $0.94 | 87.5% · $219.48 | 58.3% · $697.56 | | | 83.3% · $3.28 | 91.7% · $90.22 |
| NaturalQuestions, 2,255 q, 25% of tokens | 73.5% · $0.33 | 60.6% · $0.30 | | | 70.5% · $1.32 (1,000 q) | 52.5% · $0.87 | 73.9% · $0.30 | 78.4% · $0.65 |
| HotpotQA, 2,000 q, 40% of tokens (exact match) | 52.2% · $0.15 | 52.4% · $0.13 | | | 50.4% · $0.77 (1,000 q) | 46.7% · $0.47 | 50.5% · $0.13 | 58.0% · $0.24 |
* Dense retrieval over turns (Qwen3-Embedding-0.6B) on LoCoMo and LongMemEval; a plain reranker on NaturalQuestions
and HotpotQA. Mem0 is its top 50 memories, except top 200 on Mem0's own sample. LongLLMLingua ran on pre-registered
samples; on the same questions LLMSlim with its reranker scores 93.3% on LoCoMo (+20.0, 95% CI +15.0 to +25.0) and
72.7% on NaturalQuestions, and the default LLMSlim 53.3% on HotpotQA. A cached full history costs $4.48 per 1,000
LoCoMo questions when the cache misses.
Where LLMSlim loses:
- NaturalQuestions without the reranker. The default lexical mode scores 60.6% against 73.9% for a plain
reranker: paraphrased questions need semantic matching, so use the reranker for open-domain retrieval.
- Dense retrieval on LongMemEval. 80.6% to LLMSlim's 77.8% at 10,000 tokens and 76.1% to 74.4% at 2,500, at
about twice the cost. Neither gap is significant on 180 questions, but both point the same way. A fix that
narrowed it on development questions did not hold up on test, so it is not in 0.9.0.
- Multi-hop and temporal questions at 1,000 tokens. Mem0's top 20 memories beat the default LLMSlim on LoCoMo's
multi-hop (88.7% vs 80.3%) and temporal (93.4% vs 85.7%) questions. With its reranker LLMSlim is ahead overall
(90.5% vs 88.7%).
- A cached full history. When a conversation's questions come back to back, the cached history (95.1% for
$0.83) beats every compressed context on LoCoMo.
Long documents and agents: round 2, Core 0.8.0
Same model and prompt for every configuration; truncation gets exactly LLMSlim's token count, and every suite was
rerun once on a fresh sample after the code was frozen (benchmarks/live). 0.9.0 changes
only how plain-text chat transcripts are selected.
| Benchmark | Full context | LLMSlim 0.8 + reranker | LLMSlim 0.8 (no model) | Truncate, same budget | Tokens (full → LLMSlim) |
|---|
| LongBench v1, 210 samples (macro) | 54.9 | 54.2 | 54.9 | 45.2 | 15,532 → 5,232 |
| … fresh 210 samples | 55.1 | 55.4 | 55.0 | 46.8 | |
| LongBench v2, 180 q (accuracy) | 60.0 | 55.6 | 53.9 | 48.3 | 30,493 → 8,469 |
| SWE-bench Verified Mini, 50 issues (resolved) | 36-38 | not run | 38 | not run | 48-58k → 51k per task |
The reranker is optional (relevance="semantic", a 22M-parameter cross-encoder from the semantic extra); without
it LLMSlim runs no model at all.
NOTE
The Schema-Tax Measurement: In our v0.4.0 benchmark across 375 synthetic tool schemas in 18 catalogs written for this repository (see the dataset card), lossless schema compression saved 0 tokens (0.00%) when the baseline was already compact canonical JSON. We publish this result transparently: we never claim imaginary savings where none exist.
v0.6 Adaptive Planner (Frozen Offline Suite)
| Method | Budget success | Context reduction | Required facts | Hard constraints | Mean latency* |
|---|
| Raw full context | 71.4% | 0.0% | 100% | 100% | measurement only |
| Naive prefix | 100% | 9.2% | 100% | 100% | 2.21 ms |
| Fixed-ratio compressor | 78.6% | 9.6% | 94.0% | 96.4% | 26.55 ms |
| Adaptive planner | 96.4% | 11.4% | 100% | 100% | 208.02 ms |
The 28-case corpus covers chat, RAG, tools, mixed, external-source, and 12-language
multilingual cases. The sole adaptive infeasibility is an intentionally impossible
full-catalog case; it is reported explicitly. These task-grounded checks are not an
LLM-judge or universal quality claim. Latency is specific to the checked-in run and
hardware. See the release gate results,
raw results, and frozen
dataset. The live Sarvam harness
was not run without both explicit opt-in and credentials.
🖥️ Command-Line Interface (CLI)
Compress files, streams, or prompts directly from your terminal:
llmslim document.txt -o compressed.txt --ratio 0.5
cat input.txt | llmslim --ratio 0.4 --mode rag > rag_context.txt
llmslim document.txt --stats
📋 Changelog
Release notes for every version, including 0.12.0, are in
CHANGELOG.md and on
GitHub Releases.
Upgrading from 0.12 to 1.0: docs/UPGRADING.md lists every change that can break code, every name
that moved and where to import it from now, and how to try the release candidate (pip install "llmslim==1.0.0rc1"
once it is out).
🤝 Ecosystem Integrations
LLMSlim operates as an unopinionated context pre-processor. These integrations ship in Core; none of them sends a request for you. Every integration that spends a vendor credit is checked by a UsageGuard before the request leaves the process (per-minute rate, lifetime request cap and a unit cap in INR, USD or queries, configurable with LLMSLIM_<VENDOR>_* variables; see .env.example):
| Provider / Framework | What ships in Core | Pattern |
|---|
| Sarvam AI | llmslim[sarvam]: SarvamProvider/AsyncSarvamProvider, SarvamLanguageTools, SarvamDocumentSource, annotate_languages()/normalize_scripts(), sarvam_messages() | Local planning + official sarvamai Python SDK, guarded with dated INR prices |
| Exa | llmslim[exa]: ExaSearchSource, vendor-neutral WebSearchSource | Untrusted RAG context items with URL, date and rank provenance; spend capped in USD |
| Neo4j | llmslim[neo4j]: Neo4jContextStore | Explicit tenant-scoped temporal memory: supersede(), recall(as_of=...), history(), conflicts() |
| OpenAI | openai_compatible_request(), openai_responses_request() | Execution-free request builders; your code sends the request |
| Anthropic Claude | anthropic_messages_request() | Execution-free request builder with optional prompt-cache markers |
| Google Gemini | gemini_generate_request(), gemini_cached_content_request() | Execution-free request builders |
| vLLM / self-hosted | vllm_chat_request() | Execution-free request builder |
| OpenAI Agents SDK | llmslim[agents]: make_openai_agents_input_filter(), to_openai_agents_tools() | Text-only model-input filter; the SDK keeps execution and tool ownership |
| LangChain | llmslim[langchain]: LLMSlimCompressor | BaseDocumentCompressor for ContextualCompressionRetriever |
| LlamaIndex | llmslim[llamaindex]: LLMSlimPostprocessor | BaseNodePostprocessor for query engines |
| Haystack | llmslim[haystack]: LLMSlimDocumentCompressor | Pipeline component between retriever and prompt builder |
| gptme | llmslim[gptme]: LLMSlimContextProvider | Context provider for gptme's compaction (entry point llmslim) |
| Hermes Agent | LLMSlimContextEngine (no extra) | ContextEngine plugin; no LLM call, cache-stable prefix |
| Pydantic AI | llmslim[pydantic-ai]: llmslim_history_processor() | History processor that keeps tool calls and returns paired |
For other frameworks there is no dedicated adapter. LLMSlim works with any framework
that can pass plain message lists ({"role": ..., "content": ...} dicts) or document
strings; you write the few lines of glue that pass LLMSlim's output to your client.
📜 License & Citation
LLMSlim is open-source software licensed under the MIT License. See LICENSE for full details.
If you use LLMSlim in your research or production systems, please cite:
@software{llmslim2026,
author = {Yashvardhan Thanvi},
title = {LLMSlim: Deterministic Context Compression and Contract-Safe Tool Schemas for LLMs},
year = {2026},
publisher = {GitHub},
url = {https://github.com/Thanatos9404/llmslim}
}
Built with precision for developers everywhere. Accepted into the Sarvam Startup Program.
Agent Context Runtime (v0.7.0)
Version 0.7.0 added a per-turn context path for agent applications (0.7.0 was
not published to PyPI; install 0.7.1 or later).
The host supplies messages, retrieved documents, memories, tool results, and
authorized tool schemas; LLMSlim returns a quality-checked model input and an
explainable local trace. Model calls and tool execution stay with the host.
from llmslim import ContextRuntime, generic_request
runtime = ContextRuntime(model="sarvam-105b", max_input_tokens=8_000)
prepared = runtime.prepare_sync(
user_input="When does Acme renew?",
messages=[{"role": "system", "content": "Use verified dates."}],
documents=[{"id": "contract", "content": "Acme renews on 9 November 2026."}],
)
if prepared.feasible:
request = generic_request(prepared)
print(request["messages"])
print(prepared.trace.explain())
The runtime includes ContextEnvelope, ContextGraph, ContextPolicy,
progressive planning, quality gates, bounded in-memory sessions, and
execution-free adapters for generic, Sarvam, OpenAI-compatible, and text-only
OpenAI Agents applications. ContextTrace omits prompt bodies by default.
The existing compress() and plan_context() APIs continue to work.
See the architecture,
security model,
integrations,
migration guide,
benchmark, and
engineering report.
Cache-Aware Context Runtime (v0.7.1)
Cache planning is an opt-in layer for repeated agent turns. LLMSlim identifies
stable provider-visible prefixes, keeps append-only conversation context when
safe, fingerprints cache-sensitive settings, and reports local cache estimates
separately from provider-reported usage. A tenant-scoped CacheManager stores
metadata and provider references, never prompt bodies or KV tensors.
from llmslim import CacheManager, CachePolicy, ContextRuntime
runtime = ContextRuntime(
model="generic-128k",
cache_policy=CachePolicy(mode="auto", provider="openai", tenant_id="my-workspace"),
cache_manager=CacheManager(),
)
prepared = runtime.prepare_sync(
session_id="conversation-1",
messages=[{"role": "system", "content": "Use verified sources."}],
user_input="What changed?",
)
print(prepared.trace.to_dict()["cache"])
Execution-free request builders cover OpenAI Responses, Anthropic Messages,
Gemini cached resources, and self-hosted vLLM. An advanced in-process
Transformers adapter validates KV continuation state. OpenAI provider-stateful
continuation and Gemini cached-resource creation require explicit host opt-in.
Sarvam prompt caching remains unverified and is disabled by default. The
Context Inspector shows prefix segmentation and cache diagnostics without
calling a provider. See caching guidance,
multi-turn benchmark, and
engineering report.
Snyk provides developer security tooling for open-source projects. Learn more about its Secure Developer Program.