Obsidian vault as long-term memory cuts AI coding agent context by 67%
Someone running multiple (Claude Code, Codex, OpenCode) across several solo products built a system that splits knowledge by how long it stays useful. Project-specific facts that are true right now live in a small folder inside each project repo, and agents read that first. Durable knowledge — decisions, lessons, history, and patterns that repeat across projects — lives in a separate private , kept in local-only git with no sync and no public remote.
The vault's homepage note acts as a command center, listing open questions and what to work on next. It uses only plain Markdown and wikilinks, with no community plugins, so any agent can read it without special tooling. Retrieval follows a strict rule: agents query the vault narrowly, pulling at most 3 to 5 notes per question, and every note carries frontmatter recording its source, last-verified date, and confidence level, so is visible rather than silently trusted.
Measured in words of context loaded per session, this cut a typical session from about 7,500 words to about 2,470 — roughly a 67% reduction. The author flags two honest caveats: these dashboard-style notes need weekly upkeep or they go stale and mislead the agent, and the measurement used word count as a stand-in for actual rather than an exact figure.
Key points
- Splits knowledge into project-specific facts (small in-repo folder, read first) versus durable cross-project knowledge (separate private )
- Vault is kept in local-only git with no sync and no public remote
- A homepage 'command center' note lists open questions and next tasks for agents to check first
- Retrieval is capped at 3-5 notes per question, and every note has frontmatter with source, last-verified date, and confidence
- Reported cutting loaded context from about 7,500 to 2,470 words per session, roughly 67%, though measured in words rather than exact tokens