nanobot-runtime/skills/keep/SKILL.md

name, description

name	description
keep	Explicit immediate memory. Use when user says "keep X", "zapamatuj si X", "ulož si X", "pamatuj si X", "/keep X". Adds, deduplicates, and compacts entries in workspace/keep.md. Separate from MEMORY.md / Dream pipeline.

Keep

Explicit memory store. User says "keep X" → reformulate, write to /home/nanobot/.nanobot/workspace/keep.md, dedup, compact when too long.

File

workspace/keep.md. Flat bullet list. One entry = one line: - <terse fact>. No dates. Date adds noise, has no value for keep/discard decisions.

Write protocol

1. Extract the core fact. Drop filler ("you know", "important is that", "watch out for", "remember please").
2. Rewrite as a 5–15 word terse fact. Telegraphic style. Dashes / parentheses for context. Preserve the language of the input — never translate. Czech input → Czech entry, English input → English entry.
   • Input: "you know, Honza from marketing is allergic to peanuts"
   • Entry: Honza (marketing) — peanut allergy
   • If the entry is a decision, preference, or dead-end (not a plain fact), append the reason inline on the same line: <fact> — because <terse why>. Plain facts (allergy, deploy window, name) get no reason.
   • If it is a decision/preference/dead-end but the input gives no reason, ask the user once for the why before storing. If they supply it → append it. If they decline or it is self-evident → store without it.
3. Read workspace/keep.md (create if missing).
4. Read workspace/memory/MEMORY.md and check if a semantically similar fact already exists there (Dream may have already distilled it).
   • If similar fact in MEMORY.md → tell the user ("Already in MEMORY.md: <existing fact>. Keep anyway?") and act on their answer. Default: skip.
5. Check duplicates: case-insensitive substring match against existing lines.
   • If duplicate → ask user: replace, append as variant, or skip.
6. Append the new line.
7. If line count > 150 → run compaction (below) before responding.
8. Confirm: Kept: <terse fact>. Respond in the user's language (the model localizes the confirmation itself).

Compaction

Trigger: line count > 150, or user says "keep compact" / "udělej compaction".

1. Read full keep.md.
2. Rewrite under ~120 lines (headroom). Strategies:
   • Merge duplicates and near-duplicates.
   • Drop stale one-shot info (past meetings, transient states, expired notes).
   • Shorten verbose entries.
3. Write the new file in one go.
4. Report: Compaction: 151 → 117 lines.

Edge cases

• /keep with no content → ask "What should I remember?".
• Vague input ("remember this", "that thing") → ask for the concrete fact; do not store a placeholder.
• File missing → create it on first write.
• Multi-line input → collapse newlines to spaces; one entry = one line.

Rules

• Never store the verbatim input. Always reformulate.
• Reason (why) only for decisions / preferences / dead-ends — never for plain facts. Always terse and inline on the same line; never a separate Why: block.
• Preserve input language; never translate.
• Do not store smalltalk or meta-commentary about memory itself.
• Keep is separate from MEMORY.md and Dream. Read MEMORY.md only for the dedup check (step 4); never edit it or any Dream file from this skill.
• Touch keep.md only via this skill. Other agent paths should read it (per USER.md reference) but not edit it.