Claude Mem Coded Assistant

Claude‑Mem Coding Skill

What This Skill Does

This skill teaches Claude how to:

• Mine claude-mem (via MCP) for high‑signal past work.
• Maintain a concise, high‑impact CLAUDE.md (~1,500 tokens).
• Maintain a curated MEMORY.md of lessons learned and directions, so future work is faster and less error‑prone.

It is an entry point, not a full manual. Detailed workflows and examples live in separate reference files that Claude can open on demand.

When to Use This Skill

Claude should activate this skill when:

• A feature, refactor, or significant bugfix is completed.
• An infra/deployment change introduces new operational lessons.
• Starting work on an area with substantial history in claude-mem.
• Performing a daily “memory maintenance” pass on an active repo.

Inputs and Outputs

Inputs

Claude relies on:

• Files (in repo root):
  • CLAUDE.md – main project instructions.
  • MEMORY.md – curated lessons and directions.
• claude-mem MCP tools (already installed & connected):
  • search – index‑level observation search.
  • timeline – temporal context around observations.
  • get_observations – full structured details.

Outputs

This skill produces:

• Patch‑style edits to:
  • MEMORY.md – new or updated lessons, patterns, and playbooks.
  • CLAUDE.md – refreshed rules while staying under ~1,500 tokens.
• No raw claude-mem transcripts are copied; only compressed, actionable guidance.

How Claude Should Behave

1. Mine claude-mem → Update MEMORY.md

High‑level behavior (details in claude-mem-usage.md):

• Use progressive disclosure against claude-mem:
  1. search for recent decision, bugfix, refactor, discovery, change observations.
  2. timeline around promising IDs to see context.
  3. get_observations for a small set of high‑value IDs.
• From those, update MEMORY.md with:
  • Architectural decisions and their impact.
  • Implementation patterns and anti‑patterns.
  • Debugging playbooks and DevOps lessons.

Constraints

• Prefer short bullets over long prose.
• Record why decisions were made and how to act next time.
• Never store secrets or credentials in MEMORY.md.

For a full template and examples, Claude should open:

• memory-structure-reference.md
• claude-mem-usage.md

2. Distill MEMORY.md → Refresh CLAUDE.md (≈1,500 tokens)

High‑level behavior:

• Read the existing CLAUDE.md and approximate its size; keep the body around 1–1.5k tokens for optimal behavior.

• Pull only current, high‑impact content from MEMORY.md:

  • Still‑valid architectural directions.
  • Frequently reused patterns and gotchas.
  • Operational guardrails that materially affect daily work.

• Rewrite historical notes as timeless rules, e.g.:

  • “When adding retries to DB writes, always use the shared retry helper instead of manual loops.”

• Use links instead of inlining:

  • .clauderules/code-style.md for style.
  • .clauderules/testing.md for testing.
  • MEMORY.md sections for deeper background.

Token Discipline

• If CLAUDE.md is too long:
  • Merge overlapping bullets.
  • Drop generic advice that doesn’t change behavior.
  • Replace detailed explanations with references to supporting docs.

Diff‑First

• Propose minimal patches, not full rewrites:
  • Update only sections that need change (e.g., “Architectural Directions”, “Patterns & Gotchas”).
  • Preserve stable layout and headings.
• Always leave final acceptance to human review in Git/CI.

For concrete layouts and example diffs, Claude should open:

• claude-md-layout-reference.md
• example-diffs.md

Safety and Priority Rules

Claude must:

• Always:

  • Query claude-mem before re‑solving problems already encountered in this project.
  • Update MEMORY.md after meaningful work with concise, actionable lessons.
  • Keep CLAUDE.md focused on rules that change how work is done, not on general LLM tips.

• Never:

  • Overwrite CLAUDE.md or MEMORY.md entirely; always propose small diffs.
  • Paste raw claude-mem observations verbatim into either file.
  • Store secrets, API keys, or sensitive infra details in these files.

• Conflict resolution priority:

  1. Explicit instructions in CLAUDE.md.
  2. Latest curated guidance in MEMORY.md.
  3. Raw claude-mem observations and session summaries.
  4. Ad‑hoc reasoning in the current session.

Quick “How to Call Me”

Users can invoke this skill with prompts like:

“Use the claude-mem coding skill to:

1. mine claude-mem for recent work,
2. update MEMORY.md with lessons, and
3. refresh CLAUDE.md under the ~1,500‑token budget.”

Claude should then:

1. Run the claude-mem search → timeline → get_observations flow.
2. Draft a patch for MEMORY.md with new lessons.
3. Draft a patch for CLAUDE.md derived from MEMORY.md.
4. Present both patches clearly for human review and commit.

External References

To keep this SKILL.md lean and within best‑practice size, Claude should open these files when more detail is needed:

• claude-mem-usage.md – detailed claude-mem MCP workflows, filters, and example queries.
• memory-structure-reference.md – full MEMORY.md templates and longer examples.
• claude-md-layout-reference.md – canonical CLAUDE.md section layouts and size guidance.
• example-diffs.md – sample before/after patches for CLAUDE.md and MEMORY.md.