m2ai-portfolio Claude Memory Architect
Memory is a fingerprint. No two people should have the same one. This skill interviews you, audits existing patterns from the community, and builds you a memory system that fits your working style rather than a template.
When to Invoke
- User says "build my memory system", "design my Claude memory", "memory architect", "what should I remember between sessions"
- User notices they repeat the same context at the start of every session
- User has tried a generic memory template and it didn't stick
- User is starting a new project and wants to set up memory intentionally from the beginning
Phase 1: Interview (7 Questions, One at a Time)
Ask these sequentially. Never batch them.
- What type of work do you do most? (code, writing, analysis, client work, mixed)
- What do you most often re-explain to Claude at session start? (project context, preferences, active tasks, all of the above)
- How far back does "recent context" meaningfully go? (hours, days, weeks)
- Do you need memory across multiple projects or one project? (sets scope)
- Can you tolerate a few seconds of memory loading at session start, or does startup speed matter? (hooks vs CLAUDE.md)
- Are there things you never want Claude to remember? (explicit exclusions)
- Will you actively maintain your memory, or do you need it to decay gracefully on its own? (determines decay rules)
Phase 2: Repo Audit (Clone & Cherry-Pick)
Identify 2-3 public Claude Code memory repos relevant to the user's answers. For each:
- What is the core philosophy? (structured facts / narrative / event log / hybrid)
- What injection mechanism does it use? (CLAUDE.md / hook / external store)
- What would break if the user adopted this verbatim?
- What single pattern from this repo is worth stealing?
Summarize what to take and what to leave:
PATTERN INVENTORY
=================
Keep:
- [Pattern A] from [Repo X] — because [reason tied to user's answers]
- [Pattern B] from [Repo Y] — because [reason]
Skip:
- [Pattern C] — incompatible with [constraint user stated]
- [Pattern D] — overkill given [user's maintenance tolerance]
Phase 3: Design the Memory Spec
Build the user's personal memory spec using these building blocks. Use only the ones that earn their place.
Building Blocks
Salience — What gets written at all? Bias toward surprises, corrections, and non-obvious preferences. Skip facts Claude can re-derive from the code.
Disclosure — What can Claude surface unprompted vs only when asked? High-salience memories may surface freely. Sensitive or speculative memories stay latent.
Decay — How long before a memory becomes stale? Define a TTL per category (e.g. active tasks: 7 days, preferences: indefinite, project context: 30 days).
Promotion — What triggers a memory from short-term to long-term? (repeated relevance, user confirmation, a certain output type)
Multi-Signal Retrieval — Does retrieval rely on keyword match, semantic similarity, recency, or a combination? Pick the retrieval signal that matches the user's lookup habits.
Compaction — When the memory store grows large, what gets summarized vs discarded? Establish a compaction cadence if the user's volume warrants it.
Output a lightweight spec:
# My Memory Spec
## What I remember
- [category]: [what goes here], TTL: [decay window]
- [category]: [what goes here], TTL: [decay window]
## What I don't remember
- [explicit exclusion]: [why]
## Retrieval approach
[how memories surface — keyword / semantic / recency / explicit lookup]
## Compaction rule
[when to summarize vs delete]
Phase 4: Choose Injection Approach
Pick exactly one primary approach (or a deliberate combination):
Approach A — CLAUDE.md Entries
Write memory directly into CLAUDE.md as structured sections. Low overhead. Read at every session. Best for stable, slow-changing context (preferences, project invariants, explicit rules).
Approach B — Hooks (SessionStart / UserPromptSubmit)
A hook reads memory files at session start and injects relevant entries. Higher flexibility — memory files can be updated without touching CLAUDE.md. Best for high-volume, frequently-changing memory.
Approach C — Agent-Scoped
Memory lives in an agent's own CLAUDE.md or system prompt. Isolated from the main session. Best when multiple specialized agents need separate memory contexts.
For most users: start with Approach A for stable facts, add a SessionStart hook only if volume grows beyond what's readable inline.
Phase 5: Build and Wire
Write the initial memory content. For each entry:
## [Category]
<!-- TTL: [decay window] | Salience: [high/medium/low] -->
[memory content]
If a hook is chosen, write the hook script skeleton. If agent-scoped, identify which agent owns which memory section.
Finish with a verification test: ask Claude "who am I?" or "what are you working on?" and confirm the answer reflects the memory correctly without prompting.
Verification
- All 7 interview questions answered (no blanks)
- At least one repo was audited and pattern inventory produced
- Memory spec exists with explicit TTLs and exclusions
- Injection approach chosen with a stated reason
- Verification query produced the expected memory-grounded answer
Source
Mark Kashef — "Master ALL 7 Levels of Claude Code Memory" (2026-04-22) https://www.youtube.com/watch?v=OMkdlwZxSt8
Core technique: the 3-step clone-audit-cherry-pick process, the "memory fingerprint" philosophy (no two memory systems should look the same), and the building block vocabulary (decay, promotion, salience, disclosure, compaction, multi-signal retrieval). Injection approaches drawn from the video's three-path taxonomy.