New in liteagents: Hot Memory for AI coding tools

Most AI memory solutions focus on indexing everything — conversations, files, context — into vector databases or embeddings. But the real bottleneck was never indexing. It's retrieval. You can index 10,000 chunks and still get irrelevant results because retrieval doesn't know what matters. Claude Code does have a built-in auto-memory that loads the first 200 lines of a project memory file each session — but it writes generic learnings. What it doesn't do is systematically extract what went wrong, what you decided, or how you prefer to work. So we built a 3-step pipeline that feeds structured knowledge into that same memory slot: /stash snapshots your session context before it's lost to compaction, /friction analyzes your session logs for failure patterns and behavioral signals (turns out 58% of my sessions were "BAD" — the AI claimed success while commands were still failing), and /remember consolidates everything into a structured MEMORY.md with facts, episodes, and preferences. No database, no embeddings, just markdown that Claude reads on every session start.

The result is "hot memory" — lightweight, project-local, accumulates across sessions. It captures three types of knowledge: atomic facts (decisions, corrections), episodes (what was tried, what worked), and behavioral preferences extracted from friction analysis (like "verify exit codes before claiming done" — backed by 65+ observed failures). For context, Claude Code's instruction stack isn't a simple hierarchy — CLAUDE.md provides advisory instructions loaded per session, hooks enforce deterministic guardrails that can't be overridden, and skills/subagents/MCP tools add modular capabilities on demand. Our pipeline targets the memory layer specifically, making sure hard-won context from past sessions isn't lost and the AI doesn't repeat the same mistakes. npx liteagents to try it — works with Claude Code, Opencode, Ampcode, and Droid.