<< All versions

Skill v1.0.2

currentAutomated scan100/100
kbanc85/claudia/memory-audit

~2 modified

──Details
PublishedSeptember 15, 2026 at 05:45 AM
Content Hashsha256:b7a387c40bc15a85...
Git SHAeb63cecebc7f
Bump Typepatch
Compare with v1.0.1
──Files
Files (1 file, 4.2 KB)
SKILL.md4.2 KBactive
SKILL.md · 185 lines · 4.2 KB

version: "1.0.2" name: memory-audit description: "Show everything Claudia knows with provenance tracing and entity counts. Triggers on \"what do you know?\", \"show memories\", \"memory audit\", \"what do you remember about\". See also: memory-health for system-level stats and data quality; diagnose for connectivity troubleshooting." argument-hint: "[entity name]" effort-level: medium


Memory Audit

Show what Claudia knows. Verify claims trace to sources. Surface gaps.

Usage

  • /memory-audit -- Full system audit
  • /memory-audit [entity name] -- Audit everything about a specific person, project, or entity

Full Audit

When run without arguments, produce a system-level overview of everything in memory.

1. Summary Counts

Query the memory system for aggregate counts:

claudia memory entities search --query "*" --project-dir "$PWD"
claudia memory recall "*" --compact --limit 1 --project-dir "$PWD"
claudia memory document search --project-dir "$PWD"

Parse the JSON output from each command to get counts.

Display:

## Memory Audit - [Date]
| Category | Count |
|----------------|-------|
| Entities | N |
| Memories | N |
| Commitments | N |
| Documents | N |
| Relationships | N |

2. People (Top 10 by Importance)

claudia memory entities search --types "person" --limit 10 --project-dir "$PWD"
For each person:
claudia memory about "[person name]" --project-dir "$PWD"

Display as a table:

### People
| Name | Memories | Last Mentioned | Key Fact |
|------|----------|----------------|----------|
| ... | ... | ... | ... |

3. Projects (Top 10)

Same pattern with types=["project"]:

claudia memory entities search --types "project" --limit 10 --project-dir "$PWD"

4. Active Patterns

claudia memory session context --project-dir "$PWD"

5. Provenance Sample

Pick the 3 most recent high-importance memories and trace them:

claudia memory recall "*" --compact --limit 3 --project-dir "$PWD"
For each result, run:
claudia memory provenance trace --memory-id "[id]" --project-dir "$PWD"

Display:

### Provenance Check (3 recent memories)
**Memory:** "[content snippet]"
- Source: [episode/document/user_input]
- Document: [filename] (if linked)
- Entities: [linked entities]
- Verified: [yes/no/pending]

Entity Audit

When run with an entity name (e.g., /memory-audit Sarah Chen):

1. Profile

claudia memory about "[entity name]" --project-dir "$PWD"

Display:

## Audit: [Entity Name]
**Type:** person/project/organization
**Description:** [from entity record]
**Importance:** [score]
**First seen:** [created_at]
**Last mentioned:** [updated_at]

2. All Memories (grouped by type)

From the claudia memory about JSON response, group memories:

### Facts (N)
- [content] (importance: X, created: date)
### Commitments (N)
- [content] (importance: X, created: date)
### Observations (N)
- [content] (importance: X, created: date)

3. Relationships

### Relationships (N)
- [relationship_type] with [other_entity] (strength: X)

4. Linked Documents

claudia memory document search --entity "[entity name]" --project-dir "$PWD"

Display:

### Documents (N)
- [filename] ([source_type], [date]) - [summary snippet]

5. Provenance Chains

For each commitment or high-importance memory (importance > 0.7):

claudia memory provenance trace --memory-id "[memory ID]" --project-dir "$PWD"

Display:

### Provenance
**"[memory content]"** (commitment, importance: 0.9)
|- Source: session_summary (episode 42)
|- Episode: "Discussed Q2 goals with Sarah..."
|- Document: meeting-sarah-q2.md (transcript)
|- Verified: yes (2026-01-15)

Output Rules

  • Use the structured output format with emoji headers
  • End structured output blocks with a markdown horizontal rule
  • If the memory system is not available, say so clearly
  • Keep entity audit focused: no padding, no speculation
  • Provenance chains are the most important part: if a memory has no source, flag it

Tone

  • Factual and clean
  • Like a database report, not a narrative
  • Flag gaps honestly: "No source document linked" is useful information
← v1.0.1All versions