obsidian-agent-memory

A production-grade memory system for AI coding agents.

Stop dealing with agents that forget previous architectural decisions, hallucinate repo states, or burn tokens reading your entire codebase just to fix a typo.

Clone → Open in Obsidian → Point your AI agent → Start coding with memory.

The Problem

Karpathy's LLM Wiki pattern and projects like obsidian-wiki, second-brain, and llm-wiki-compiler are excellent for general knowledge management. But they're designed for knowledge ingestion — not for the tight feedback loop of active software development, where:

• Token budgets matter — agents read everything, even when they only need one fact.
• Vaults bloat — agents over-update notes, introducing drift and noise.
• Repo truth gets ignored — a vault note says 416 tests, but the repo now has 500. The agent trusts the vault and hallucinates.

The Solution

Obsidian Agent Memory introduces three patterns that fix this:

🧊 Context Capsules

Dense, 60-line-max notes that compress operational knowledge into retrievable boundaries. Instead of reading a 300-line architecture doc, your agent reads capsule-database.md and knows exactly what ORM you use, where the schema lives, and what breaks if you touch it.

📶 Tiered Retrieval Protocol

A cost-conscious reading order: capsules first, full docs only if needed. Most coding tasks resolve at Tier 2 — agents stop escalating as soon as they have the answer.