Office Hours III: Backing Up Claude Chats with Obsidian, Bridging Chat and Cowork, What Fable Is For, and the Context-Rot Problem

Akhil's answer is the Obsidian pattern: a local markdown vault that Claude and Claude Code write into, so every conversation and every generated document lives on your disk with graph links between them. Search 'Claude with Obsidian' on YouTube for the setup. It is configurable to include attachments. Two caveats he adds: this works with Claude because Claude has a local app to integrate - ChatGPT and Gemini do not (Sarah confirms they need manual export) - and artifacts inside a Claude project at least warn you before deletion.

There is no native shared memory between chat and Cowork (or Claude Code), and no project-inside-a-project. The workarounds: tell Claude explicitly to 'refer to the chat that mentions X' so it pulls a summary across, or - Swaminathan's follow-up - generate a complete markdown summary of the chat before moving into Cowork, accepting that some detail is lost to the context window either way.

Fable's /loop and /goal modes are the 'auto research' pattern Martin raised: the model orchestrates its own research loop toward a goal. Akhil relays the official guidance not to over-constrain it with a skills file - it has its own agentic loop - and his own workflow: use Fable once to scaffold the big thing, then downgrade to Opus or Sonnet for iteration. Non-coding uses: synthesis over large datasets, research and white-paper writing, and cybersecurity or app auditing. For Paul's construction spec-book app - ingest the whole book, answer questions, and flag where things could go wrong - his 'uninsured opinion' is that Opus-class handles it unless the job needs long autonomous reasoning.

Akhil frames the ceiling as physical - context windows are bounded by hardware - and predicts specialised domain agents (Harvey for law) orchestrated by a master agent. Today's mitigations: Claude Code's documented sub-agents feature, scoping a task to a cheaper model (UI polish to a small model, core logic to the strong one); an external memory layer; and the research terms to read - 'agent bootstrapping' and 'context hydration'. For the participant whose camera-and-voice agent forgets everything on wake-up and drowns in daily markdown logs, the answer is RAG for fast retrieval plus explicit state management plus long-term memory, combined, not any one of them.

Akhil endorses the migration without hesitation: platform-locked storage (Bolt's) becomes a pricing and lock-in risk at scale; Supabase, a single-purpose database company, has less incentive to hike prices than a platform that sells you everything. Keep Claude Code connected to GitHub and treat the database as yours. He cannot diagnose the tokens-scale-with-database symptom without detail, but the principle stands: your data, your provider.

The participant wanted a Hermes agent on a Hostinger VPS running 24/7 against a local Gemma. Akhil's fix: the agent lives on the VPS, but the model should be an API call to OpenRouter - a marketplace of cheap open models (a DeepSeek price is quoted loosely) - so nothing depends on a home machine being awake. Related: Claude Code now supports local models via Ollama, and OpenCode is the bring-your-own-model alternative if it does not.

Either run open-source models on servers you control so nothing leaves the building, or use the providers' documented guardrail and anonymisation features so PII (account holders, transaction details) is redacted before processing. Which one is a cost-and-risk decision, not a technical one.