claude-code - 💡(How to fix) Fix [Feature Request] Persistent project memory — sessions forget everything on close, forcing users to keep many sessions open

Problem

Claude Code has no persistent project-level memory. When a session is closed, all conversation context — architectural decisions, debugging history, what was tried and why, in-progress reasoning — is lost entirely.

claude --resume only recovers a compressed summary, which loses critical details. CLAUDE.md and the memory system help with rules and preferences, but cannot replace the rich, fine-grained understanding that builds up during a working session.

The real-world consequence

As a heavy user (10+ active projects, daily multi-hour sessions), I am forced to keep 10-15 Claude Code sessions open simultaneously just to preserve context. Closing any session means the next one starts significantly "dumber" — it forgets decisions we made, problems we solved, and the current state of work.

This directly causes another reported bug: keeping many sessions open means many Node.js processes with their TUI render loops and timers all competing for CPU and terminal I/O, which contributes to the input freeze issue reported in #62999 (and historically in #35903, #58498, #31194).

These two problems are linked:

• No persistent memory → users keep many sessions alive → TUI freezes more often
• TUI freezes → users lose input and have to re-type → frustration compounds

What I need

A persistent, per-project memory that survives session close/restart. When I open Claude Code in a project directory, it should remember:

1. What we were working on — not just a summary, but the specific decisions, trade-offs, and reasoning
2. What was tried and failed — so it doesn't suggest the same broken approaches again
3. The current state of progress — which tasks are done, which are pending, what's blocked
4. Project-specific knowledge — architecture patterns, naming conventions, gotchas discovered during work

The existing memory system (MEMORY.md + memory files) is a step in the right direction but is too manual and limited. It stores preferences and facts, not the evolving working context of a session.

Proposed solution

• Auto-persist session context per project directory, not just compressed summaries
• Rich resume that restores the full working understanding, not a lossy summary
• Project knowledge graph that accumulates across sessions — every session makes the next one smarter, not equal or dumber

Why this matters commercially

Heavy API users like myself spend significant money on Claude Code. The value proposition breaks down when:

• I pay for 8 hours of context-building in a session
• Close the session
• Next session starts from near-zero, requiring me to re-pay for the same context-building

Persistent memory would make every dollar spent compound in value rather than evaporate on session close.

Environment

• Claude Code 2.1.109
• 10-15 concurrent projects (Odoo 18, FastAPI, Docker, various AI integrations)
• Ubuntu 24.04, Linux 6.8.0-117-generic
• Daily heavy usage, Anthropic API (Max plan)