The Problem Is Bigger Than You Think
You've been working with Claude Code for hours. You've explained your architecture, fixed bugs together, made important decisions. Then suddenly... Claude has no idea what you're talking about.
It's not just new sessions. Claude loses context in multiple ways:
- New sessions — Start fresh, explain everything again
- Context compaction — Long conversations get summarized, details lost
- Token limits — Big codebases push out earlier context
- Weeks of work — That bug fix from 3 weeks ago? Gone
Every time this happens, you're re-explaining the same architecture, the same decisions, the same gotchas. Over and over.
The Solution: External Memory
I built ContextVault — an external context management system for Claude Code. Instead of relying on Claude's internal context (which gets lost), your knowledge lives in markdown files that Claude reads automatically.
How It Works
- You work with Claude — fixing bugs, making decisions, building features
-
ContextVault captures knowledge — using 25 slash commands like
/ctx-error,/ctx-decision,/ctx-doc -
Knowledge persists externally — in
.claude/vault/as markdown files - Every session, Claude reads your vault — full context, instantly
# Install in 30 seconds
curl -sL https://ctx-vault.com/install | bash
Why External Memory Beats Internal Context
| Problem | Internal Context | ContextVault |
|---|---|---|
| New session | ❌ Lost | ✅ Preserved |
| Context compaction | ❌ Summarized away | ✅ Preserved |
| Token limits | ❌ Pushed out | ✅ Preserved |
| Weeks later | ❌ Forgotten | ✅ Preserved |
| Team sharing | ❌ Impossible | ✅ Commit to git |
Key Features
Two-Tier Vault System
-
Global vault (
~/.claude/vault/) — Cross-project knowledge, patterns, best practices -
Project vault (
./.claude/vault/) — Project-specific architecture, decisions, configs
25 Slash Commands
| Command | Use When |
|---|---|
/ctx-error |
Fixed a bug |
/ctx-decision |
Made an architecture choice |
/ctx-doc |
Learned something useful |
/ctx-plan |
Working on multi-step task |
/ctx-handoff |
Ending your session |
/ctx-bootstrap |
New project — auto-document codebase |
Smart Hooks
ContextVault reminds you to document at meaningful milestones — after completing tasks, fixing bugs, or ending sessions. Configurable enforcement levels: light, balanced, or strict.
Real Example
Before ContextVault (week 3 of a project):
You: "Fix the auth bug"
Claude: "What auth system are you using? Can you explain the architecture?"
You: *sighs, explains for the 10th time*
After ContextVault (week 3 of a project):
You: "Fix the auth bug"
Claude: "I see from P001_auth.md that you're using JWT with 15min expiry
and Redis for refresh tokens. Checking P005_error_jwt.md for recent issues..."
Getting Started
# 1. Install globally
curl -sL https://ctx-vault.com/install | bash
# 2. Initialize in your project
/ctx-init
# 3. Auto-document your codebase
/ctx-bootstrap
# 4. Work normally — ContextVault captures knowledge
# 5. Next session — Claude remembers everything
Why I Built This
I was tired of:
- Re-explaining my project architecture every few days
- Losing important bug fix context after long conversations
- Watching Claude forget decisions we made together
Inspired by concepts from recursive language models (arxiv:2512.24601), I wanted Claude to have persistent memory that survives everything — new sessions, compaction, token limits, time.
ContextVault is:
- 100% local — Your data stays on your machine
- 100% free — No subscriptions, no cloud fees
- Open source — MIT license, fork it, modify it
- Simple — Just markdown files
- Portable — Commit your vault to git, share with team
Not Like Other Solutions
Unlike paid "AI memory" services:
- No monthly fees — Free forever
- No cloud storage — Your data never leaves your machine
- No vendor lock-in — It's just markdown files
- No privacy concerns — 100% local
Links
- Website: ctx-vault.com
- GitHub: github.com/ahmadzein/ContextVault
- Docs: ctx-vault.com/docs
If you use Claude Code, give ContextVault a try. Your future self will thank you when Claude remembers that obscure bug fix from three weeks ago — even after context compaction, token limits, and new sessions.
Star on GitHub if you find it useful! ⭐
What's your biggest frustration with AI context loss? Let me know in the comments!
Top comments (0)