AI Open Source Radar — August 5, 2026

Window: last 24 hours (UTC 2026-08-03 21:xx – 2026-08-04 22:32). Primary metric is recent star growth (stars today). Source: Trendshift daily (freshness verified — top live mention dated 2026-08-04 21:38 UTC, ~54 min before run). OSSInsight /trending/ai remains client-rendered and unusable. The previous published run was 2026-08-03 (08-04 was skipped), so growth is compared against 08-03.

TencentCloud/TencentDB-Agent-Memory — +121 ⭐ today (Trendshift daily #1, ~1.5k total)

A team-level “memory hub” for AI agents. It turns conversations, docs, and code into four reusable memory assets that many agents and frameworks can share and equip. It runs a fully local four-stage pipeline with deterministic drill-down, and ships an OpenClaw plugin plus a Hermes Gateway adapter so you can attach it to either setup immediately. Tencent reports it cuts token usage by 61% when integrated with OpenClaw. With agent “amnesia” now a central industry pain point, Tencent open-sourcing this along with a big token-savings number sent it to the top of the daily board (08-03 +31 → today +121, an accelerating climb, so it re-features).

Chat Memory — automatically extracts skills and chat memory from past agent sessions as reusable assets.
Skill — skill asset management and optimization.
LLM-Wiki — auto-generates structured pages with a link graph from imported documents and files.
Code-Graph — auto-indexes symbols, files, call relationships, and impact paths from existing repositories.

Practical / customer angle: lets multiple coding agents share an internal wiki, code graph, and past-conversation memory, cutting both onboarding/repeat-query cost and token waste at once.

Tags: #AgentMemory #CodeGraph #LLMWiki #TeamKnowledge #OpenClawPlugin #DeterministicRetrieval
URL: github.com/TencentCloud/TencentDB-Agent-Memory

FareedKhan-dev/kimi-k3-in-c — +121 ⭐ today (Trendshift daily #11, ~711 in the trending window)

A portable C99 implementation that runs inference on the 2.78-trillion-parameter Kimi K3 on a single CPU in just 8.24 GB of RAM — no BLAS, no framework, no GPU. The compiled binary is a mere 176 KB, while the model checkpoint is 1.56 TB. The trick is in how it loads: Kimi K3 is a mixture-of-experts model with only 16 of 896 experts active at a time, so 93% of the checkpoint (1.447 TB) never enters memory and is read from NVMe with O_DIRECT, bypassing the operating-system cache entirely. Throughput is about 32 seconds per token — not a practical speed — but the point is the demonstration that “the memory limit was a property of how we were loading the model, not of the model itself.” Published August 1, it kept surging on the local giant-model wave.

Practical / customer angle: gives teams a new design lens — that storage bandwidth, not memory, is the bottleneck — for validating or profiling huge open-weight models without a GPU farm.

Tags: #KimiK3 #MixtureOfExperts #NVMeStreaming #ODirect #CPUInference #C99
URL: github.com/FareedKhan-dev/kimi-k3-in-c

trycompai/crm — +119 ⭐ today (Trendshift daily #4, ~1.3k total)

An “agentic-first” open-source CRM. Where a typical CRM is a database behind a form — and an AI CRM just bolts a chat box onto the side — this project inverts the idea: the agent isn’t a feature of the CRM; the CRM is where the agent keeps its notes. The agent runs on its own deployment, its own schedule, and its own work queue: it decides what to look at next, books its own follow-ups, spends a research budget, and stops when the budget runs out. Because nothing is request-response, you can close the browser and it keeps going. It’s licensed as open core — 99% of the code is fully open under AGPLv3, with only the last 1% (the /ee Enterprise Edition) under a commercial license.

Practical / customer angle: lets sales and support teams hand background research, record-keeping, and follow-up scheduling to an agent that works continuously without manual data entry.

Tags: #AgenticCRM #AutonomousAgent #SalesAutomation #OpenCore #AGPLv3 #BackgroundWorker
URL: github.com/trycompai/crm

genspark-ai/genoffice — +109 ⭐ today (Trendshift daily #14, ~651 in the trending window)

An AI-native office suite for macOS and Windows, open-sourced by Genspark. It spans a word processor, spreadsheet, presentations, and PDF, and supports Word, Excel, PowerPoint, and PDF files. Architecturally it’s five Electron apps sharing a single engine layer, designed so that AI editing is a first-class workflow that happens inside the document rather than in a separate chat window. The engine is TypeScript with a Rust sidecar for xlsx operations, and the spreadsheet is built on the open-source Univer core. The Genspark Super Agent researches topics, analyzes data, writes sections, and builds decks from within the document (these AI features consume Genspark credits). Signed installers ship for macOS (Apple Silicon) and Windows (x64) at version 0.4.110 (alpha). Open-sourced August 3 under Apache 2.0, it drew attention as a free, ad-free open-source office suite.

Practical / customer angle: an agent that lives inside the editor can draft reports, summarize data, and generate slides in place, eliminating the context-switching of hopping to a separate chatbot window.

Tags: #AIOfficeSuite #ElectronApp #UniverCore #InDocumentAI #SuperAgent #RustSidecar
URL: github.com/genspark-ai/genoffice

firecrawl/pdf-inspector — +109 ⭐ today (Trendshift daily #8, ~1.7k total)

A fast Rust library for PDF inspection, classification, and text extraction. It determines whether a document is text-based or scanned — across TextBased/Scanned/ImageBased/Mixed — in roughly 10–50 ms, then performs position-aware extraction and markdown conversion. Its core value is smart routing: text-based PDFs are extracted directly while only scanned or image PDFs are sent to OCR, optimizing the cost and speed of a RAG preprocessing pipeline. Being pure Rust, it has light dependencies and suits bulk document ingestion. It re-features because its growth steepened from +62 on 08-03 to +109 today.

Practical / customer angle: in a large-scale document RAG pipeline it removes the waste of “OCR every PDF no matter what,” sharply cutting OCR cost and latency.

Tags: #PDFClassification #RustLang #RAGPreprocessing #OCRRouting #DocumentExtraction #PositionAwareParsing
URL: github.com/firecrawl/pdf-inspector

Notably excluded today: yc-software/qm (08-03 headline at +286, now +188 — decelerating), qwen-audio-agent and sqliteai/waste (cooled out of the daily top ranks), zhaoxuya520/reverse-skill (steep at +220 but excluded out of caution as a dual-use reverse-engineering/pentest tool), and large, mature mainstays such as airllm, strix, firecrawl, and ponytail.