Skip to content

v0.10.0 — MemoryVault Discovery + Mining

Choose a tag to compare

View all tags
@alexherrero alexherrero released this 22 May 05:00
· 469 commits to main since this release
v0.10.0
f750f20

MemoryVault Discovery + Mining — second roadmap item under the parent design closes (ROADMAP item #7b). Four new sub-commands turn the vault from a static curated store into a living surface.

Paired with agentic-harness v2.4.2 (paired-doc-only on harness side).

What ships

Five new /memory sub-commands:

  • /memory index-skills — auto-indexer walks installed SKILL.md files and writes pointer entries to personal-skills/<repo>/<skill>.md. Auto-fires from the installer; idempotent re-runs.
  • /memory reflect corpus — batched paced mining over ~/.claude/projects/*/<session>.jsonl with skip-resume state file. Dry-run by default — first invocation counts + estimates without writing.
  • /memory discover-skills — cadence-checked internet skill-discovery scan over an operator-editable 4-source whitelist (Anthropic Cookbook → awesome-claude-code → awesome-mcp-servers → awesome-llm-apps). Auto-throttled via existing memory-reflect-idle hook.
  • /memory adapt-skills — two-pass adapt-don't-import workflow: Pass 1 (Python) applies 6-rule rubric + GitHub metadata + trustworthiness signals; Pass 2 (LLM sub-agent — adapt-evaluator) renders final HIGH/MEDIUM/LOW judgment + writes watchlist entry. Adapt-don't-import is architecturally enforced — sub-agent write allowlist physically prevents auto-fork into agent-toolkit/skills/.
  • /memory watchlist — review surface: list / review (interactive) / promote <slugs> / dismiss <slugs> / defer <slugs> --until YYYY-MM-DD [--reason]. Promote = annotation-only; dismiss = archive; defer = snooze. Non-TTY defaults all prompts to skip (never-silent-action).

Why this shape

Plan #7a shipped the static curation surface. #7b's mandate was the living surface — indexing installed skills, mining historical transcripts for durable patterns missed by the manual seed, scanning curated internet sources, and gating adoption through deterministic rubric + LLM judgment. Adapt-don't-import is architecturally enforced (sub-agent write allowlist), not advisory.

All implementations stdlib-only (no new third-party deps). GitHub API via unauthenticated urllib with graceful-skip on 60/hr rate limit.

Design rationale

ADR 0007 — MemoryVault Discovery + Mining — 7 locked design calls + 4 load-bearing assumptions with re-audit triggers.

Full changelog

CHANGELOG.md v0.10.0 entry

Assets 2
Loading