Originally created by: Thanatos9404
Before: the startup-programs banner, nine large badges, "The context layer for production AI agents" with a line about compression, then a compress() quickstart. Memory, MCP and the client setups started about 560 lines down.
After:
Memory("memory.db"), add, search(..., budget=1500)) and TypeScript (npm install llmslim against llmslim serve). A tip for clients that time out on the first uvx download. One plain star line.uvx, tiktoken's tokenizer file, the semantic model) and the lexical-matching trade-off.Nothing below was deleted. The banner moved to "Startup programs". The tagline, the Core description and the other badges (Python versions, tests, coverage, Live Studio, Sarvam) moved to "LLMSlim Core and LLMSlim Platform". "Try it in 60 seconds" is now "Compress a prompt in 60 seconds", and its star line was merged into the new one. No accuracy, cost or saving claims and no comparisons were added.
docs/assets/llmslim-memory-demo.svg (188 KB, animated, about 30 s). It is a Python session against llmslim 0.11.1 installed from PyPI in a fresh venv on Linux (WSL):
search("Am I allergic to anything?", budget=50) returns only the peanut-allergy turns;search("Where did I move to?", budget=50) returns only the message added live.docs/assets/demo/memory_session.py seeds the SQLite file and types the lines into a real code.InteractiveConsole, so only the typing speed is scripted and every output line comes from llmslim. It was recorded with asciinema rec (asciinema 2.4.0 from PyPI); the raw recording is docs/assets/demo/memory-demo.cast. The SVG was rendered with svg-term-cli@2.1.1 (npm). The rebuild commands are in the script's docstring.
examples/, each with a README showing its real output)| Example | What it runs | Tested by |
|---|---|---|
claude-code-memory |
Claude Code setup, plus session.py: an MCP client starts llmslim mcp over stdio, calls remember in one session, then recall and list_memories from a new server process |
tests/test_examples.py |
langchain-agent-memory |
LangChain 1.x create_agent with remember/recall tools on Memory and LLMSlimToolOutputMiddleware, on a scripted fake chat model |
tests/test_examples.py |
pydantic-ai-memory |
Pydantic AI agent with memory tools in deps and llmslim_history_processor, on FunctionModel |
tests/test_examples.py |
fit-context-big-logs |
fit_text (the function behind fit_context and /v1/context/fit) on a generated CI log or a file |
tests/test_examples.py |
typescript-client |
the published npm llmslim@0.1.0 against llmslim serve: add, search, context.fit |
npm test in the TypeScript client CI job |
examples/README.md indexes these and the existing scripts. tests/test_examples.py also runs memory_quickstart.py. It starts all examples as separate processes at once, so it adds about 12 to 15 s locally, and it skips the examples whose framework is missing (for example on Python 3.9). The CI change is one step in the existing TypeScript client job (npm ci && npm test in the example folder).
claude mcp add ... -- <command>, Codex codex mcp add <name> -- <command> and Copilot CLI copilot mcp add <name> -- <command>: checked against each client's docs. The exact uvx --from "llmslim[mcp]" llmslim mcp command was run against PyPI 0.11.1 with an MCP client (server 0.11.1, five tools, remember then recall).gemini extensions install <repo url>: checked against its docs. Their extensions page now says Gemini CLI was replaced by Antigravity CLI on 18 June 2026, so this line may need a follow-up.hermes plugins install llmslim --enable: llmslim.yaml is in hermes-agent's plugin-catalog/ (NousResearch/hermes-agent#142212, merged), and the docs list --enable.[#86] adds two lines under "In Claude, Cursor or any MCP client" (after the Gemini CLI line) and an entry under Unreleased/Added in CHANGELOG.md. This PR leaves that section as it is: the one-click links are repeated in the new first screen, not moved. Its CHANGELOG entries are under Unreleased/Documentation, so the two should merge cleanly. Once [#86] is merged and the extension is published, the "Use it in 30 seconds" table could get a VS Code extension row.
The ... [N lines omitted] marker from fit_text/compress on line-structured text counts a collapsed run of similar lines as one line. In the fit-context example, [2 lines omitted] stands for about 250 omitted test lines, and [74 lines omitted] for 192. The cause is in llmslim/focus.py around line 930: hidden run members are not added to gap. The example README shows the real output as printed.
ruff check . and ruff format --check . pass, with ruff 0.15.21 and 0.16.10.pytest tests/test_claude_plugin.py tests/test_examples.py tests/test_mcp_server.py tests/test_memory.py: 69 passed, 2 skipped (no LangChain or Pydantic AI in that venv).tests/test_examples.py: 7 passed in venvs with every framework, on Python 3.12 (pydantic-ai-slim 2.60, langchain 1.4.7) and 3.10 (pydantic-ai-slim 2.54). The Pydantic AI example also ran on 1.107.7.npm test in examples/typescript-client: passes against llmslim serve.
Originally posted by: vercel[bot]
The latest updates on your projects. Learn more about Vercel for GitHub.