Antialzheimer

Anti-Alzheimer - Memory Health & Cognitive Architecture Maintenance

Version: 0.1

Why this exists

AI agent memory systems degrade silently. Memories freeze at write-time while rules evolve. Projects complete but their memories persist. Cross-references break. Contradictions accumulate. The result: an agent that advises with full confidence but partial context.

This skill detects and repairs that drift.

Operation Routing

User action / argument	Read file	Modifies files?
/antialzheimer or check (default)	op-health-check.md	No - read-only scan
full or consolidate	op-full-consolidation.md	Yes - backup first
session or review	op-post-session.md	Maybe - proposes, user approves

Cognitive Layer Model

Understanding where information belongs is the core skill. Wrong placement = wrong weight.

Layer	Loaded	Weight	Content type	Examples
CLAUDE.md (global)	Always, all projects	Highest	Cross-project rules	Language, tone, safety
CLAUDE.md (project)	Always, this project	High	Project identity, rules	Stack, conventions, delegation
.claude/rules/	Always, this project	High	Operational patterns	Learned patterns, debugging, docs index
MEMORY.md (index)	Always	Medium-High	Instructional descriptors	One-line actionable pointers per memory
Memory files	On-demand	Medium	User context, project state	Profile, strategies, active projects
docs/system/	On-demand	Lower	Session artifacts	SCRATCHPAD, CHANGELOG, DECISIONS

Key insight: every line added to an always-loaded layer reduces adherence to ALL other lines in that layer. The ROI of pruning is high. Memories in files are suggestions; rules in CLAUDE.md are enforced.

What to detect

Issue	Severity	Description
Silent contradiction	HIGH	Memory says X, CLAUDE.md says NOT-X. Agent receives conflicting instructions
Zombie memory	MEDIUM	Memory about completed project or obsolete context still occupying cognitive space
Orphan reference	MEDIUM	Memory references file/memory that no longer exists
Missing cross-reference	LOW	Related memories don't reference each other, risking "partial personality" advice
Stale descriptor	LOW	MEMORY.md descriptor doesn't match actual file content
Budget overrun	MEDIUM	Always-loaded tokens significantly exceed recommended limits
Graduation candidate	LOW	SCRATCHPAD pattern repeated 3+ times, should promote to learned-patterns
Premature promotion	MEDIUM	Rule in CLAUDE.md that was promoted without sufficient evidence (< 3 observations)

What NOT to save as memory

Before creating or keeping a memory, apply this filter:

• Code patterns, architecture, file paths -> derivable from codebase. Don't save.
• Git history, recent changes -> git log / git blame are authoritative. Don't save.
• Debugging solutions -> the fix is in the code, the context in the commit message. Don't save.
• Anything already in CLAUDE.md or rules files -> duplicate. Don't save.
• Ephemeral task state -> use tasks/plans for current conversation. Don't save.

Memory is for what is NOT derivable: user profile, strategic context, organizational knowledge, calibration feedback, project state that transcends code.

Graduation Pipeline

The lifecycle of an observation becoming a rule:

Session error/discovery
    -> SCRATCHPAD.md (raw, timestamped)
        -> 3+ repetitions OR critical impact
            -> learned-patterns.md (graduated pattern)
                -> Repeated across projects OR universal applicability
                    -> CLAUDE.md (permanent rule)

Each promotion increases enforcement weight but consumes instruction budget. Promote deliberately.

Failure Modes

These are the most common ways memory consolidation goes wrong:

1. Silent contradiction - A memory freezes a rule at write-time. The rule evolves in CLAUDE.md. Now the agent gets conflicting instructions. Detection: diff memory content against CLAUDE.md rules + learned-patterns.

2. Partial personality - Incomplete memory sets generate confident but incomplete advice. Origin: this was discovered when two PCs had disjoint memory sets - one knew HOW to work (process), the other knew WHY (strategy). Neither had the full decision chain.

3. Memory hoarding - Saving everything as memory when most of it is derivable from code, git, or docs. This bloats the system and increases noise. Test: "Could I derive this by reading the codebase?"

4. Premature rule promotion - Moving a single observation straight to CLAUDE.md. Every line in CLAUDE.md reduces adherence to all other lines (linear degradation for frontier models). Promote only after 3+ observations or critical impact.

5. Over-consolidation - Merging so much context into one memory that it loses specificity. Every memory should have a clear "when to use" - if it's always relevant, it might belong in CLAUDE.md instead.

6. Zombie accumulation - Completed projects, obsolete contexts, resolved decisions lingering as active memories. They compete for attention with current priorities.