Feature request: Native persistent memory / hooks or custom MCP support for stateless sessions

Hi Kimi Code CLI team,

I'd like to share a gap I’ve noticed while using Kimi Code CLI with long-running projects.

Problem
Each session is currently stateless: the agent cannot remember decisions, lessons, or project-specific notes from previous sessions. When I return to a project, I have to re-explain context, re-state conventions, and re-learn from past mistakes.

My setup
I run a local persistent-memory layer (claude-mem + graphify + markdown decision/lesson records + Ollama bge-m3 embeddings) that works great with Claude Code and Codex. It exposes read-only tools like search_memory, get_record, and prompt-recall over a local daemon.

Requested capabilities

1. Lifecycle hooks(SessionStart, UserPromptSubmit, etc.) so a local script can inject a recall block or trigger memory extraction, or
2. Custom MCP server / skill registration so users can wire their own local memory daemon into the agent context, or
3. A built-in persistent memory layer that can query an external/local memory store before each prompt.

Benefits

• Cross-session decision continuity
• Fewer repeated mistakes
• More context-aware architectural decisions
• Parity with Claude Code / Codex ecosystems

Thanks for considering this.

No files or tests are named. Start by tracing the CLI session lifecycle around SessionStart and UserPromptSubmit, along with custom MCP or skill registration paths; the issue needs a decision on which capability to implement. Done should mean a local memory daemon can provide context across sessions through the selected integration.