PMB Local-First Memory for AI Coding Agents

功能对比

Direct Answer

PMB local-first memory is useful when AI coding agents need to remember project facts across sessions without sending every detail to a hosted memory service. Treat it as a local recall layer, not a replacement for committed repo instructions.

PMB Setup Checklist

Start with a minimal local memory scope. The first setup should remember stable engineering facts, not private customer data or temporary task notes.

• Create a local workspace for the repository and name the project clearly.
• Store durable facts such as package manager, test commands, architecture boundaries, and recurring review preferences.
• Keep secrets, tokens, production logs, customer data, and private messages out of long-term memory.
• Connect one coding tool first, then test whether retrieved memory improves a real bug fix or refactor.
• Add an owner and review cadence so stale memory can be pruned.

Memory seed:
project = ai-coding-tools-guide
package_manager = npm
build = npm run build
seo_check = npm run check:seo
safe_default = inspect files before editing; report skipped checks

PMB Vs CLAUDE.md Vs Repo Memory

Use each memory layer for a different job. CLAUDE.md and AGENTS.md should describe what the repo expects. PMB-style memory should remember useful context that changes less often than a chat but may not belong in committed files.

• Put team-approved commands and rules in CLAUDE.md or AGENTS.md.
• Put editor-specific globs and patterns in Cursor Rules.
• Put reusable but personal workflow context in local-first memory.
• Put product decisions and architecture references in docs that humans can review.

Claude Code, Codex, Cursor, And Zed

A local memory layer should adapt to the way each coding tool starts work. Claude Code benefits from explicit project memory and terminal commands, Codex benefits from portable task context, Cursor benefits from rules near the edited files, and Zed-style editor workflows benefit from concise project summaries.

• Claude Code: keep CLAUDE.md authoritative and let local memory add recurring repo facts.
• Codex: mirror stable constraints into AGENTS.md so delegated tasks start with portable instructions.
• Cursor: keep .cursor/rules focused on file-level editing behavior and use memory for broader project preferences.
• Zed or other editors: keep memory summaries short enough to paste or retrieve without flooding the prompt.

Local SQLite Memory Pattern

SQLite is a practical local-first store because it is file-based, inspectable, and easy to back up. Keep the schema boring so developers can audit what the agent remembers.

tables:
  projects(id, name, repo_path, created_at)
  memories(id, project_id, scope, summary, source, updated_at)
  reviews(id, memory_id, action, reviewer, reviewed_at)

scopes:
  command | architecture | style | testing | release | personal

Security And Privacy Checklist

Long-term memory changes the trust boundary. Review memory writes the same way you review agent tool access.

• Do not store API keys, session cookies, tokens, passwords, or raw credentials.
• Do not store private customer data, production incidents, or proprietary logs unless the team has approved that local retention.
• Keep memory files outside public repos unless they are intentionally committed project instructions.
• Add a delete and review workflow before sharing memory across machines.
• Prefer summaries with source pointers over pasted raw files.

常见问题