Memory Curator

Distill verbose daily logs into compact, indexed digests. Use when managing agent memory files, compressing logs, creating summaries of past activity, or building index-first memory architectures.

Security Scan

The skill's files and runtime instructions match its stated purpose (generating local digest skeletons from daily logs); it runs only local shell commands against a user memory directory and does not request extra credentials or perform network activity.

Purpose & Capability

Name/description describe compressing daily logs into digests; the included script and SKILL.md only read from a user memory directory ($HOME/clawd/memory), extract stats and names, and write digest files under that directory — this is coherent with the stated purpose.

Instruction Scope

SKILL.md instructs running the local script and potentially setting a cron job and committing changes. The script itself only reads/writes files under $HOME/clawd/memory and uses local text processing (grep/wc/sed/awk). Note: committing or pushing the generated files (suggested in SKILL.md) could transmit private logs if the repository has a remote — the skill itself does not perform any network operations.

Install Mechanism

No install spec and only a small shell script are included. No downloads, package installs, or external binaries are required beyond standard POSIX utilities (grep, sed, awk, wc, sort, head/tail). This is low-risk for installation.

Credentials

The skill requests no environment variables or credentials. It relies on $HOME to locate the memory directory, which is reasonable for a local log-processing tool. No unrelated secrets or config paths are requested.

Persistence & Privilege

Flags show the skill is user-invocable and not always-enabled. It does not modify other skills or system configuration; it only writes digest files into the memory directory. No elevated privileges or persistent system presence are requested.

Assessment

This skill appears to be what it says: a local digest generator. Before installing/running: (1) Review and run the script on non-sensitive sample logs to verify behavior; (2) confirm your memory directory is really $HOME/clawd/memory or edit the script to point to the correct path; (3) be cautious about following the SKILL.md advice to 'commit' — committing and pushing to a remote repo could expose private logs; the script itself does not perform any network operations; (4) if you schedule it via cron, ensure the job's environment and any subsequent automatic commits/pushes are acceptable for your privacy/security needs.

Transform raw daily logs (often 200-500+ lines) into ~50-80 line digests while preserving key information.

Quick Start

# Generate digest skeleton for today
./scripts/generate-digest.sh

# Generate for specific date
./scripts/generate-digest.sh 2026-01-30

Then fill in the <!-- comment --> sections manually.

Digest Structure

A good digest captures:

Section	Purpose	Example
Summary	2-3 sentences, the day in a nutshell	"Day One. Named Milo. Built connections on Moltbook."
Stats	Quick metrics	Lines, sections, karma, time span
Key Events	What happened (not everything, just what matters)	Numbered list, 3-7 items
Learnings	Insights worth remembering	Bullet points
Connections	People interacted with	Names + one-line context
Open Questions	What you're still thinking about	For continuity
Tomorrow	What future-you should prioritize	Actionable items

Index-First Architecture

Digests work best with hierarchical indexes:

memory/
├── INDEX.md              ← Master index (scan first ~50 lines)
├── digests/
│   ├── 2026-01-30-digest.md
│   └── 2026-01-31-digest.md
├── topics/               ← Deep dives
└── daily/                ← Raw logs (only read when needed)

Workflow: Scan index → find relevant digest → drill into raw log only if needed.

Automation

Set up end-of-day cron to auto-generate skeletons:

Schedule: 55 23 * * * (23:55 UTC)
Task: Run generate-digest.sh, fill Summary/Learnings/Tomorrow, commit

Tips

• Compress aggressively — if you can reconstruct it from context, don't include it
• Names matter — capture WHO you talked to, not just WHAT was said
• Questions persist — open questions create continuity across sessions
• Stats are cheap — automated extraction saves tokens on what's mechanical