Today's digest

# memvid/memvid **URL:** https://github.com/memvid/memvid **One-liner:** A single-file, serverless memory layer for AI agents that replaces complex RAG pipelines with fast, persistent, and portable memory. **Relevance to oss-digest:** high (85/100) **Integration:** depend-on-it ## Summary Memvid is a serverless memory layer for AI agents that provides instant retrieval and long-term memory via a single file. ## Why it's useful here oss-digest uses DeepSeek to generate daily digests of new open-source projects; it needs memory to avoid re-processing duplicates and to maintain conversational context across sessions. Currently likely uses ad-hoc storage; Memvid's portable .mv2 capsules could replace this with versioned, crash-safe memory. ## Suggested use Integrate the Node.js SDK (npm @memvid/sdk) into the digest generation pipeline to store and recall already-seen projects, and to provide the AI with persistent context across daily runs. ## Novelty / why now Novel concept of 'Smart Frames' inspired by video encoding, enabling append-only, immutable memory capsules with time-travel debugging and sub-5ms recall, all in a single file. ## Risks Young project (350 days); core in Rust but SDKs abstract this; single-maintainer risk despite 24 contributors; potential API instability before v1. ## Safety scan - Risk level: **low** - Stars: 15479 (age 350d, 44.23 stars/day) - Last push: 6 days ago - Contributors: 24 - License: Apache-2.0 - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes Apache-2.0 license, low risk, no postinstall hooks or suspicious patterns; 24 contributors and active development.

# rohitg00/agentmemory **URL:** https://github.com/rohitg00/agentmemory **One-liner:** Agentmemory provides persistent memory for AI coding agents via MCP, hooks, and a REST API, with confidence scoring, knowledge graphs, and hybrid search. **Relevance to oss-digest:** high (85/100) **Integration:** depend-on-it ## Summary Persistent memory for AI coding agents that enables agents to remember across sessions with confidence scoring and knowledge graphs. ## Why it's useful here oss-digest's AI agent currently runs DeepSeek analyses without persistent memory; integrating agentmemory would allow it to remember past digests, avoid re-analyzing the same repo, and build a knowledge graph of topics and trends over time. ## Suggested use Import agentmemory as an MCP server or use its npm library to store and retrieve analysis results, confidence scores, and relationships between repos. ## Novelty / why now Combines Karpathy's LLM Wiki pattern with production-grade features (confidence scoring, lifecycle, knowledge graphs) and zero external database dependencies. ## Risks Very new repo (77 days) with aggressive star growth; single maintainer (rohitg00); may have unstable API or future breaking changes; verify compatibility with your Next.js version. ## Safety scan - Risk level: **low** - Stars: 6575 (age 77d, 85.39 stars/day) - Last push: 0 days ago - Contributors: 13 - License: Apache-2.0 - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes Low risk - no suspicious patterns, no postinstall hooks, Apache-2.0 license. However, the repo is very new (77 days) with rapid star growth (6.5k), which could indicate hype; evaluate stability and long-term maintenance.

# BenedictKing/ccx **URL:** https://github.com/BenedictKing/ccx **One-liner:** Go-based multi-provider AI API proxy with web admin, channel orchestration, failover, and key management. **Relevance to oss-digest:** high (85/100) **Integration:** depend-on-it ## Summary Unified AI API proxy supporting Claude, OpenAI, Gemini, and Codex with built-in web admin, failover, and key rotation. ## Why it's useful here oss-digest uses DeepSeek via OpenAI-compatible API. ccx can proxy DeepSeek (via OpenAI endpoint) and add failover, multi-key management, and monitoring. Currently keys are likely hardcoded. ## Suggested use Deploy ccx as sidecar proxy; point oss-digest's AI calls to ccx's /v1/chat/completions endpoint. Use ADMIN_ACCESS_KEY for web admin. ## Novelty / why now Not novel; similar to LiteLLM/OpenRouter but with integrated UI and dual-key auth. ## Risks Young project (102 days), single-maintainer risk despite 11 contributors, recently spiked stars (possible hype). Requires managing a Go binary. ## Safety scan - Risk level: **low** - Stars: 603 (age 102d, 5.91 stars/day) - Last push: 0 days ago - Contributors: 11 - License: MIT - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes MIT license, no suspicious patterns, 11 contributors, moderate stars spike (603 in 102 days).

# Tencent/WeKnora **URL:** https://github.com/Tencent/WeKnora **One-liner:** Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki. **Relevance to oss-digest:** medium (65/100) **Integration:** depend-on-it ## Summary An LLM-powered knowledge platform that ingests documents, builds RAG, and auto-generates a wiki with agent capabilities. ## Why it's useful here oss-digest already pulls OSS projects and uses DeepSeek for triage. WeKnora could index collected project info into a searchable knowledge base with agent-driven summarization and cross-linking. ## Suggested use Run WeKnora as a sidecar service, use its API to ingest curated OSS project metadata, then replace current DB queries with WeKnora's RAG and wiki mode. ## Novelty / why now Combines RAG, ReAct agent, and auto-wiki generation with multi-source ingestion (Feishu, Notion, etc.) and 20+ LLM providers. Active development by Tencent. ## Risks License ambiguity (NOASSERTION vs MIT), large Go codebase, requires external vector DB, active development may cause breaking changes. ## Safety scan - Risk level: **low** - Stars: 14825 (age 295d, 50.25 stars/day) - Last push: 0 days ago - Contributors: 85 - License: NOASSERTION - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes License unclear (NOASSERTION but MIT badge in README). Requires significant infrastructure. Not audited. May have telemetry.

# supertone-inc/supertonic **URL:** https://github.com/supertone-inc/supertonic **One-liner:** Lightning-fast on-device multilingual TTS using ONNX, with bindings for Python, Node.js, Swift, Rust, etc. **Relevance to oss-digest:** medium (65/100) **Integration:** cherry-pick ## Summary On-device multilingual TTS using ONNX, with Node.js support. ## Why it's useful here oss-digest produces daily digests of open-source news; adding TTS would let users listen to the digest, increasing engagement and accessibility. The Node.js SDK can be integrated into Next.js API routes to generate audio for each digest item. ## Suggested use Use supertonic's Node.js SDK to generate audio files for digest items, embed an audio player in the UI. Consider pre-generating audio during digest creation and storing in S3 or similar. ## Novelty / why now On-device TTS supporting 31 languages, optimized for edge inference, with a Voice Builder feature. ## Risks Node.js binding may not be production-ready; ONNX runtime native dependency may not work in serverless environments. Large model downloads (Git LFS) require caching strategy. Project primarily Swift-based; Node.js path is an example, not official SDK. ## Safety scan - Risk level: **low** - Stars: 3769 (age 176d, 21.41 stars/day) - Last push: 6 days ago - Contributors: 4 - License: MIT - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes Low risk: MIT license, no suspicious patterns, active development. However, Node.js binding is example-grade; production readiness unclear. Model downloads are large and require Git LFS. ONNX runtime must be available in deployment environment.

# statewright/statewright **URL:** https://github.com/statewright/statewright **One-liner:** State machine guardrails for AI coding agents, constraining tool access per workflow phase. **Relevance to oss-digest:** medium (65/100) **Integration:** cleanroom-rebuild ## Summary State machine guardrails that control which tools your AI agent can use in each phase. ## Why it's useful here oss-digest uses a two-stage DeepSeek pipeline to generate digests; statewright could constrain the LLM's tool usage (read-only during planning, write-only during generation) to reduce flailing and improve output quality. ## Suggested use Study statewright's state definitions and transition guards, then cleanroom-rebuild a similar concept in Python/Next for oss-digest's agent loop. ## Novelty / why now Repackages classic state machines as a deterministic Rust engine + MCP plugin to enforce per-phase tool restrictions on AI agents. ## Risks Single-maintainer, no license, unproven at scale; rebuilding in Python avoids Rust compilation dependency. ## Safety scan - Risk level: **medium** - Stars: 188 (age 9d, 20.89 stars/day) - Last push: 0 days ago - Contributors: 1 - License: unknown - Postinstall hooks: none - Suspicious patterns: none - Notes: single-contributor repo with notable stars ### Reviewer safety notes Single-maintainer repo (<9 days old, 188 stars) with no license; rapid star growth may be inorganic; risk of abandonment.

# MemoriLabs/Memori **URL:** https://github.com/MemoriLabs/Memori **One-liner:** Memori is an LLM-agnostic memory layer that persists agent execution and conversation state, with both TypeScript and Python SDKs. **Relevance to oss-digest:** medium (55/100) **Integration:** cherry-pick ## Summary Agent-native memory infrastructure for persistent state. ## Why it's useful here oss-digest uses a two-stage DeepSeek pipeline; Memori can remember which projects the user has already seen or engaged with, improving triage and personalization. ## Suggested use Register Memori with the LLM client to maintain a memory of previously digested projects, user preferences, and feedback to refine future digests. ## Novelty / why now Strong LoCoMo benchmark results (81.95% accuracy at 5% of full-context tokens) and both cloud and BYODB options. ## Risks Same cloud dependency; also may conflict with existing memory approach. ## Safety scan - Risk level: **low** - Stars: 14416 (age 293d, 49.20 stars/day) - Last push: 0 days ago - Contributors: 34 - License: NOASSERTION - Postinstall hooks: none - Suspicious patterns: none - Notes: (none) ### Reviewer safety notes License is Apache 2.0, no postinstall hooks, no secrets, low risk. However, default usage depends on Memori Cloud (SaaS) which may raise data privacy concerns. BYODB mitigates this.