Project Sessions Analyzer

A cross-session memory consolidation skill for Claude Code — the local, human-governed version of "dreaming." From any working folder, it finds the linked Claude Code project under ~/.claude/projects/, reads all of that project's session transcripts, detects patterns that are only visible across sessions (not within one), and reconciles them into the project's Auto Memory folder.

Why

Reflecting after a single session is provisional — it can mistake a one-off for a rule (false positive) or miss a real pattern that didn't look like one yet (false negative). Reading all sessions at once is the corroboration gate:

• 1 session supports a learning → status: provisional (kept, but not eligible to freeze)
• ≥2 sessions corroborate → status: confirmed (eligible to freeze into a skill)
• later sessions contradict an existing memory → corrected or removed (false positive)
• recurring pattern that was never written → added (false negative)

It only writes to the project's Auto Memory folder. It never edits a skill folder.

What it does

1. Locate + link — resolves the CWD to its project folder; creates + links an Auto Memory folder if missing.
2. Extract signal — compact per-session evidence (prompts, corrections, tool failures, denied permissions, tool usage) instead of raw transcripts.
3. Cluster across sessions — group evidence into candidate patterns with a corroboration count.
4. Reconcile memory — confirm / add / correct / keep-provisional / log skill-gap, respecting the project's own memory conventions.
5. Apply + report — writes the change-set (or --dry-run to preview) and reports exactly what changed.

Install

npx skills add juststartingtocode/skill-folder-project-sessions-analyzer -g --all

Use

In Claude Code, from a working folder with past sessions:

"Analyze my sessions and update auto memory." "What patterns show up across my sessions?" "Run the project sessions analyzer."

Needs ≥2 sessions to confirm anything; with one session it still scaffolds memory and writes provisional notes (and says so).

Scripts (deterministic, stdlib-only, Windows-safe)

• project-sessions-analyzer/scripts/locate.py — discovery, linking, memory scaffold → JSON manifest.
• project-sessions-analyzer/scripts/extract-signal.py — compact cross-session evidence.

The judgment (clustering + reconciliation) lives in SKILL.md; the scripts handle the mechanical parts.