docs: memex agent-memory pointer in the global summary (personal tooling)

[vpciii]

Wires the memex MCP memory server into your personal global summary — the tool-appropriate home — while keeping methodology.md tool-agnostic and untouched.

What it adds (to templates/global-CLAUDE.md, clearly labeled "Personal tooling — not the shared methodology"):

• Search project memory (memory_search / memory_get_project) at the start of substantial work.
• Write/summarize (memory_write / memory_summarize_and_store) at the end.
• Guardrail: memex is recall, not the record — durable decisions stay in the repo artifacts (ADRs, glossary, specs, architecture.md); it never becomes a second un-versioned home for them (ADR 0022).

Why here and not the methodology: memex is a specific tool, and the methodology is deliberately tool-agnostic ("document the artifact, not the tool" / "no premature tooling commitments"). CLAUDE.md is the methodology's designated home for tool-pointers; the global summary is your personal one. No adoption impact — consuming projects copy project-CLAUDE.md, not the global summary.

Post-merge: deploy templates/global-CLAUDE.md → ~/.claude/CLAUDE.md (the per-machine step, ADR 0018) so your live global memory picks it up.

🤖 Generated with Claude Code

[claude]

Claude finished @vpciii's task in 1m 9s —— View job

---

Review

• Read PR diff
• Read governing artifacts (ADR 0018, ADR 0022, global-CLAUDE.md, CHANGELOG conventions)
• Review against project standards

Judged against this repo's own artifacts, this is a clean, well-reasoned change. The "why here, not the methodology" argument holds up:

• Consistent with ADR 0018. templates/global-CLAUDE.md is the canonical, repo-tracked source of the always-loaded summary; editing it here (with the per-machine ~/.claude/CLAUDE.md copy as the post-merge release step) is exactly the workflow that ADR prescribes. The same-PR honesty rule is satisfied — the CHANGELOG.md entry rides along in the same diff.
• Doesn't violate the tool-agnostic stance. methodology.md is untouched; the new section is fenced off as "Not part of the shared methodology … drop it if you copy the methodology," and the guardrail ("memex is recall, not the record") correctly defers durable decisions to the versioned artifacts per ADR 0022. No conflict with single-source-of-truth.
• Conventional Commits + adoption labeling. docs: is the right type; "Adoption: none" is honestly stated.

Two minor, non-blocking observations:

1. CHANGELOG adoption vocabulary. The changelog preamble (lines 7–9) defines adoption impact as one of two categories — reference-only or per-project action. This entry introduces a third, Adoption: none. It's the accurate label (this genuinely isn't a methodology change, so neither named category fits), but it's slightly outside the stated convention. Fine as-is; flagging only in case you'd rather the preamble acknowledge a "none" case so the vocabulary stays closed.

2. The file's charter stretches a little. The header comment on global-CLAUDE.md and ADR 0018 frame the file as "the always-loaded summary of this methodology" — edited in the same PR as the methodology change it summarizes. This section is personal tooling that summarizes nothing in methodology.md. The PR justifies it well (the file doubles as the deployed personal ~/.claude/CLAUDE.md, so it's the natural home), and the explicit "drop it if you copy" labeling keeps it from leaking into adopters. Defensible — just noting the file is now slightly more than a pure methodology summary.

Neither is worth holding the PR. No correctness, security, or process-drift issues found.

One small consistency nit if you touch it again: the first bullet writes the fully-qualified mcp__memex__memory_search but then memory_get_project / memory_write unprefixed. Reads fine (prefix established once); purely cosmetic.