INFUSE vs Obsidian

INFUSE and Demandbase Announce Partnership to Turn Account Intelligence into Pipeline

INFUSE and Demandbase partner to turn account intelligence into pipeline across the whole buying group.

Robust Nightly LLM Log Drain Survives Sleep and Races

A technical walkthrough of a macOS shell script (vault-auto-ingest.sh) that nightly ingests Claude conversation logs into a local Obsidian vault, commits them to a private Git repo, and generates a daily brief. The author describes three real failure modes that drove the final design: sleep-related freezes when the Mac sleeps, double-execution races between scheduled and manual runs (noted on 2026-06-10), and a service-side timeout/outage that caused uniform slot failures on 2026-06-13. The resilient design uses four launchd slots (4:55, 8:20, 10:45, 12:15), caffeinate -i -s, atomic lockdir via mkdir, idempotent step markers for partial retry, explicit timeouts with --kill-after, preflight checks (network, TCC/Full Disk Access), and deterministic overwrites for LLM-generated numbers.

• The author built a macOS shell script to drain Claude conversation logs into an Obsidian vault, commit to a private repo, and leave a daily briefing on the desktop.
• Three real failure incidents drove the design: a double-execution race on 2026-06-10, a service-wide timeout where every slot timed out on 2026-06-13, and repeated sleep-related freezes.
• The script runs in four launchd slots at 04:55, 08:20, 10:45, and 12:15 and uses per-step markers (STEP2A/STEP2B) to allow idempotent partial retries.

LLM Wiki: Protecting Human-Refined Notes with Curated Layer

The article describes agnosticBrain, a GitHub implementation that extends Andrej Karpathy's LLM Wiki pattern by adding explicit ownership layers to protect human-refined knowledge from being overwritten by LLM ingests. agnosticBrain defines directory-level ownership rules — including a read-only curated/ folder for human-canonical notes — and introduces workflows (/curate, /propose) so the LLM can propose diffs but never modify curated content without explicit human approval. The project documents nine kernel operations (AGENTS.md) and provides a proposal, logging, and index-based pipeline to close the learning loop while preserving manual edits.

• Andrej Karpathy published his LLM Wiki gist in April 2026 proposing a compiled, persistent wiki maintained by an LLM.
• agnosticBrain (GitHub: Macorreag/agnosticBrain) extends the LLM Wiki pattern by introducing an explicit curated/ read-only layer to protect human-refined notes.
• agnosticBrain defines four directories with strict ownership rules: raw/ (human immutable), wiki/ (LLM-writable), curated/ (human-canonical, LLM never modifies), and archivo/ (system archive).

AI-native Second Brain Using Multi-RAG and Knowledge Graphs

The article outlines a proposed architecture for an AI-native "Second Brain": a persistent knowledge layer that combines multiple retrieval strategies (Multi-RAG), knowledge graphs, long-term memory, and an MCP API to connect LLMs to structured organizational knowledge. Instead of relying solely on vector search, the design integrates semantic search, keyword/full-text search, knowledge-graph queries, metadata filtering, and reranking. Sources such as Obsidian notes, GitHub, Slack, databases, and documents are ingested, processed into a knowledge layer (vector index, search index, graph store) and exposed via a retrieval engine and MCP so LLMs (e.g., Claude, GPT) can both read context and write back structured memories. The goal is a Knowledge OS that preserves knowledge independently of any single model and provides richer evidence+relationship context for reasoning.

• The author proposes an AI-native "Second Brain" that combines Multi-RAG, knowledge graphs, long-term memory, and MCP to provide unified knowledge to LLMs.
• Multi-RAG combines multiple retrieval methods: semantic (vector) search, full-text/keyword search, knowledge graph queries, metadata filtering, and reranking.
• The architecture ingests sources including Obsidian (notes), GitHub (code/PRs), Slack (conversations), databases, and documents into a knowledge layer with a vector search index, search index, and graph store.