Feature Request: Persistent cross-session project memory

Problem

Claude Code starts every session completely cold. There is zero continuity between sessions — no memory of what was worked on, what decisions were made, what the current state of a project is.

For users managing long-running, complex projects (multiple services, bots, deployments), this creates a significant and recurring productivity bottleneck:

• Every session begins with re-establishing context through documentation files
• Users must manually re-explain project architecture, conventions, recent changes
• Time spent re-contextualizing scales with project complexity, not with the task at hand
• The assistant behaves like a brilliant contractor who forgets everything overnight

Current workarounds (all insufficient)

1. CLAUDE.md — static documentation file loaded at session start. Must be manually maintained. Doesn't capture dynamic state (what changed today, what's in progress, what failed).
2. Manual memory plugins (e.g., claude-mem) — require explicit queries at session start. Entirely dependent on user discipline. If the user doesn't prompt it, nothing is loaded.
3. Verbose prompting — users paste context into every session. Wastes tokens, wastes time, scales poorly.

None of these approximate what a senior developer naturally has: memory of what they did yesterday.

What's being requested

Native persistent project memory that survives across sessions:

• Key facts about the project (architecture, services, conventions)
• Recent decisions and their rationale
• Current state of in-progress work
• What was changed in the last N sessions
• Accumulated preferences and feedback from the user

Proposed solution (minimal viable)

1. Auto-summarize at session end: When a session ends, generate a compact summary of what changed, what's in progress, and any important decisions made.
2. Auto-load at session start: On the next session start, surface that summary as initial context — similar to how a developer reads yesterday's notes before starting.
3. Project-scoped: Memory should be scoped to the working directory / project, not global.

Real-world impact

In a production environment with ~10 active services (Telegram bots, APIs, databases, deployment pipelines), re-establishing context takes 5–15 minutes per session. For a user working in short focused sessions throughout the day, this compounds into hours of lost productivity per week.

The core value proposition of an AI coding assistant is multiplicative productivity. Cross-session amnesia turns it into a tool that's brilliant in isolation but exhausting to use continuously.

This is table stakes for any tool that aspires to be a long-term development partner, not just a one-shot assistant.

---
Submitted by a user running Claude Code continuously on a production VPS, managing 10+ active services.