Skip to content

skills.md

Latest commit

 

History

History
74 lines (59 loc) · 3.09 KB

skills.md

File metadata and controls

74 lines (59 loc) · 3.09 KB
Raw

Skills — Skill = Memory (D2.1-D2.4)

Skills are agent-read Markdown pages: the durable, distilled instructions an agent accumulates ("how we deploy", "what the user prefers in code review"). No embeddings for retrieval — the agent decides what to load via progressive disclosure: wiki_list → wiki_search → wiki_read.

Skills are wiki pages of the first-class skill type — plain Markdown with frontmatter, living on disk (git-versionable), indexed in FTS5.

The lifecycle

DREAM: skill: <distilled instruction>        (agent writes it mid-conversation)
        │  C1.12 auto-save marker
        ▼
L3 episode tagged dream_skill                 (importance 0.95, review-gated)
        │  nightly phase 4: skill_promotion   or memory_skill_promote (manual)
        ▼
skill wiki page (Markdown, ≤4KB)              wiki_read → progressive disclosure
        │  nightly phase 6: skill_reinforce   (reads → importance +0.05)
        │  promotion merges into existing pages by title/topic-prefix
        ▼
shared SSOT  ~/skills-ssot/skill/*.md         scripts/sync_skills.py → all agents

Tools

Tool Purpose
wiki_list(wiki_type="skill") the disclosure index — titles + path
wiki_search(query) titles + tags + 200-char snippets
wiki_read(path) full page content (the read leg)
wiki_add(wiki_type="skill", ...) author/update a skill directly
memory_skill_promote(...) promote episodes (episode_ids) or write an agent-distilled skill (title+content)

memory_skill_promote is idempotent: promoted episodes get the skill_promoted tag and are skipped on retry. Promotion merges into an existing skill page when the episode title matches (full title or the topic prefix before :) instead of forking a duplicate.

Lint

skill_too_large — a skill page over 4KB is flagged (concept cap: API table at the top, <100 lines). Split skills instead of growing them.

Shared SSOT across agents (D2.3)

The write surface is a git repo; agents read synced copies:

python scripts/sync_skills.py --bootstrap   # create ~/skills-ssot (git repo)
# edit / commit skills in ~/skills-ssot/skill/*.md
python scripts/sync_skills.py               # sync into all live agent wikis
  • Sync uses WikiManager.sync_external: copy + sha256 dedup + index — each agent's SQLite index stays consistent (symlinks were rejected for this).
  • Each agent syncs in its own subprocess (MCP_MEMORY_DATA_DIR isolation); aiosqlite connections are closed before exit (no shutdown hang).
  • Targets: ~/.mcp-ariel-memory-{hermes,mimocode,cowagent}. Add an agent by extending the AGENTS map in scripts/sync_skills.py.
  • Cron-ready: run the bare script on a schedule; hash-dedup makes repeats no-ops.

Where things live

Artifact Path
Skill pages (per agent) ~/.mcp-ariel-memory-<agent>/wiki/user/skill/*.md
Shared SSOT ~/skills-ssot/skill/*.md (git)
Sync script scripts/sync_skills.py
Promotion pipeline features/skill_pipeline.py
Nightly phases 4/6 hooks/user_hooks.py::_nightly