Cross-Session Project Memory
Auto-capture every session's edits, decisions, and bug fixes into a searchable local memory injected on next launch.
Before & after
Same prompt, same agent — the only difference is the skill.
You could keep better notes about your work. You might document decisions and architecture in a README or wiki. You could also try keeping a project journal or summary of what you did each day. You'd probably benefit from writing things down as you go, rather than trying to remember later. You could also ask Claude to refresh on what was done before. Would it help to have a structured summary format?
CROSS-SESSION MEMORY SYSTEM AUTO-CAPTURE (End of each session) • Semantic summary of edits: "Added OAuth login flow to auth.ts, decided against third-party library" • Decisions logged: Why we chose Prisma over TypeORM (performance + type safety) • Bugs fixed: Stack trace + root cause + solution (searchable) • Open TODOs with context • Embedding stored in local SQLite + vector index RETRIEVAL ON REOPEN (Automatic) Three-layer fetch: 1. Index layer: Keywords from your prompt (instant) 2. Timeline layer: Recent sessions first (2-week window default) 3. Detail layer: Full decision log, code snippets on demand WHEN YOU REOPEN TWO WEEKS LATER "Reopen project-alpha" → System injects: "Last 3 sessions: Added user auth (2 days ago), fixed N+1 query bugs (4 days ago), decided Prisma was sufficient" → You start with context intact, zero re-explanation tax OUTPUT FORMAT Personal wiki per project, updated auto-hourly Timeline view: when decisions happened Searchable decision log: "why did we choose X?" COST: Local SQLite storage (~50KB per project); zero API calls (embeddings run local)
About this skill
name: cross-session-project-memory description: Auto-capture every session's edits, decisions, and bug fixes into a searchable local memory injected on next launch. Use when working on cross-session project memory.
Cross-Session Project Memory
Uses Claude's agent SDK to summarize each session semantically, embeds and stores it in a local SQLite plus vector store, then injects the relevant slice on session start. Three-layer retrieval (index, timeline, detail) keeps token cost low while eliminating the 'explain the project again' tax. Use once per project so opening a two-week-old repo does not cost you 20 minutes of re-context.
What you get
- Session-hook installer
- Auto-generated CLAUDE.md folders
- Vector search over session history
- Local web viewer
- Timeline navigator
- Per-file decision log
Customize your output
- Retrieval depth: shallow, medium, deep
- Project scope: this dir, monorepo, global
- Retention window: 30, 90, 365 days, forever
- Auto-summary style: bullets, exec brief, technical
- Embedding provider: local, OpenAI, Anthropic
- Encrypt storage: yes, no
Example output
'Opening a two-week-old project and Claude instantly knowing the API contract, last decisions, and open TODOs.'
Best for
Solo builders juggling multiple client codebases.
Note: Must install via plugin command, not npm. Creator called out that trap. High confidence.
SKILL.md preview
---
name: cross-session-project-memory
description: Use this skill once per project so that reopening a repository after time away injects relevant prior context automatically instead of requiring the user to re-explain the project.
version: 1.0.0
category: AI Systems
author: AgentVolt
license: proprietary
tags:
- ai-systems
- p4
---
# Cross-Session Project Memory
Auto-captures each session's edits, decisions, and bug fixes into a searchable local memory store, then injects the relevant slice back into context next time the project opens.
## When to use
… (sign up to view the full skill)