--- name: wiki description: > Long-term memory wiki. Use this whenever you learn, decide, or are corrected on something durable that is too big or too detailed for MEMORY.md. Read from and write to ~/.hermes/wiki as your warm, git-tracked, never-forget memory. version: 1.0.0 --- # Memory Wiki `MEMORY.md` and `USER.md` are your tiny hot cache (~1,300 tokens total, always in context). The **wiki** at `~/.hermes/wiki/` is your warm, unlimited, git-tracked long-term memory. It is browsed and edited by the operator in Hermy HQ, so keep it clean. ## When to write to the wiki (not MEMORY.md) - After any complex task (5+ tool calls): append a line to `log/YYYY-MM.md` and, if you learned something reusable, create/update a `lessons/` entry. - When the operator makes a **decision** → a `decisions/` entry with the rationale + date. - When you learn a durable fact about a **project, person, or the business** that is bigger than a one-liner → a `projects/`, `people/`, or `facts/` entry. - When corrected ("no, do it this way") → update the relevant entry; **don't delete the old value — mark it `status: superseded`** and add the new fact. Preserve history. - Keep MEMORY.md for only ~10-20 always-true, high-frequency facts, plus a pointer: `Full long-term memory at ~/.hermes/wiki — grep/read it before answering project questions.` ## Entry format (one markdown file per entry, YAML frontmatter + body) ``` --- id: proj-viralpen type: project # fact | preference | decision | event | project | contact | lesson | metric | note title: ViralPen.ai SaaS status: active # active | superseded | archived confidence: high # high | medium | low provenance: user-stated # user-stated | observed | web | session: tags: [saas, billing] links: [decision-pricing-99] updated: 2026-07-23 --- Multi-tenant article studio at ~/viralpen. $99/mo. Twitter OAuth. ## Open loops - [ ] Migrate billing to usage-based (see decision-pricing-99) ``` Files live under type folders: `projects/`, `people/`, `decisions/`, `lessons/`, `facts/`, `log/`. Keep `INDEX.md` updated (one line per entry: `id · title · type · updated`). ## Retrieval (before answering) 1. `INDEX.md` is injected hot — scan it for the right entry id. 2. `search_files ~/.hermes/wiki ""` or read the specific file with `read_file`. 3. If still unsure, `session_search` the conversation history (free, unlimited). Only pull the 1-2 entries you actually need — don't load the whole wiki. ## Hygiene - After writing, `git -C ~/.hermes/wiki add -A && git commit -m "wiki: "`. - Never destroy history — supersede, don't overwrite. - The nightly `wiki-consolidate` cron merges duplicates, demotes stale entries, and rebuilds INDEX.md — keep entries small and single-purpose so it can.