Our support bot has no memory, on purpose — a five-file agent architecture
Pitiful-Surround-285 · reddit · 2026-09-07
The author shipped a support agent for Orca built from just five Markdown files (one SKILL.md plus four reference docs) — no code, no vector store, and deliberately no memory.
Why no memory
- Freshness: they ship daily; fields and limits change, so a bot that remembers last week's answer is confidently wrong this week. It fetches the live docs page each turn and cites the URL — "its knowledge is a map, not a memory."
- Stranger logic: memory suits agents with ongoing jobs; a support bot talks to a different stranger each session, so nothing one person asked should shape the next answer. It gets no memory tools at all.
Tools are equally minimal: exactly three — web search, web extract, and time. Introspection is excluded because a public support endpoint has no business reading its own tenant.
The five files: SKILL.md (identity, hard rules: never invent specifics, always cite, never touch credentials); answer-protocol.md (classify into nine question types, resolve to a docs page, fetch, answer, cite); docs-map.md (question-to-page mapping; never cite the marketing site for limits/prices/endpoints); answer-rules.md (answer in the first sentence, use the words on the page, no sales pitch); unknowns.md (five kinds of not knowing and what a good "I don't know" contains).
More from coding & agent
- Building a 'meta harness' to unlock lab subscription subsidies—and a market trading privacy for pricing — sethlazar · 2026-09-07
- Building a neutral Codex-class harness requires running the original inner harness, dev says — joshalbrecht · 2026-09-07
- Dev Builds Timestamped Summary Site for All MLOps Community Talks, With an MCP Server — strickvl · 2026-09-07
- Shipping an Obsidian plugin with GPT-6 Astra: scan limits, SVG icons, review delays — vista8 · 2026-09-07
- Grok Build 1.0.22 ships subagent messaging, first-party MCP server and safer Auto mode — mark_k · 2026-09-07
- Agents on a 16GB MacBook Air M5: Qwen 3.5-9B 4-bit can't yet build a working app — Fluid-Author-9566 · 2026-09-07