ai-memory AGENTS.md

AI Agent Instructions

This file is for AI coding assistants (Cursor, Claude Code, Codex, GitHub Copilot) working on the ai-memory codebase. Read this first before touching code.

Project Overview

ai-memory is a CLI tool + MCP server that extracts structured knowledge from AI editor conversations and saves it as git-trackable Markdown files. Targets Node.js >= 18.

• npm package name: ai-memory-cli (DO NOT rename — see docs/decisions/2026-04-24-naming.md)
• Binary name: ai-memory
• GitHub owner: hyxnj666-creator
• Current version: 2.5.0 (in tree; 2.4.0 last published on npm — see ROADMAP.md and CHANGELOG.md)
• v2.5 features in tree: ai-memory try (no-key demo), rules --target skills (Anthropic Skills), --redact at LLM call sites, Codex CLI as 5th editor, CCEB v1.1 F1 64.1%, LongMemEval-50 adapter, README "1M-context FAQ", dedup quality improvements
• v2.6 features in tree (also not yet on npm): ai-memory link (memory↔commit linking, weighted Jaccard scorer, <!--links> frontmatter), init --schedule (cross-platform daily cron: launchd/crontab/schtasks), single-chunk dedup fix + cross-type TODO subsumption (targets F1 75%+), dashboard graph enhancements (edge type differentiation, type filter toggles, hover highlight). Test suite 585.
• Before publishing v2.5.0: read docs/v2.5-maintainer-handoff.md — two maintainer-only tasks remain (v2.5-03 marketplace submissions + v2.5-07 AGENTS.md downstream eval).
• Runtime deps: @modelcontextprotocol/sdk (for serve) and zod (for bundle import validation). NO other runtime deps.

Architecture

src/
├── index.ts              # CLI entry, sets emitWarning filter, dispatches to commands
├── cli.ts                # Argument parser (manual, no library)
├── types.ts              # ALL shared types live here
├── config.ts             # Loads .ai-memory/.config.json
├── public.ts             # Library entry (exports for programmatic use)
├── commands/             # One file per CLI command (14 commands)
│   ├── extract.ts        # Orchestrates extraction pipeline
│   ├── list.ts           # Lists conversations with status
│   ├── search.ts         # Keyword + hybrid search
│   ├── recall.ts         # Git-history time-travel retrieval (v2.4) — uses git/log-reader
│   ├── rules.ts          # Exports rules: Cursor .mdc and/or AGENTS.md (--target, v2.4)
│   ├── resolve.ts        # Marks memories resolved/active
│   ├── summary.ts        # LLM-generated project summary (supports --source-id / --convo)
│   ├── context.ts        # Generates continuation prompt (supports --source-id / --convo)
│   ├── init.ts           # Project initialization
│   ├── reindex.ts        # Rebuild embeddings; --dedup retroactive cleanup (v2.2)
│   ├── watch.ts          # Auto-extract on conversation changes
│   ├── dashboard.ts      # Launch local web UI
│   ├── export.ts         # Portable JSON bundle (v2.3)
│   ├── import.ts         # Import portable JSON bundle (v2.3)
│   ├── doctor.ts         # Health check (runtime/editors/LLM/store/embeddings/MCP, v2.4)
│   ├── try.ts            # No-API-key demo: bundled scenario → AGENTS.md inline (v2.5-02)
│   └── link.ts           # Scan git commits → auto-link to memories (weighted Jaccard, v2.6)
├── sources/              # Conversation parsers (one per editor)
│   ├── cursor.ts         # ~/.cursor/projects/*/agent-transcripts/
│   ├── claude-code.ts    # ~/.claude/projects/*/*.jsonl
│   ├── windsurf.ts       # Windsurf state.vscdb (SQLite via node:sqlite, Node 22+)
│   ├── copilot.ts        # VS Code chatSessions/*.json
│   └── detector.ts       # Auto-detects available sources
├── extractor/
│   ├── ai-extractor.ts   # Chunking, LLM calls, dedup, quality filter
│   ├── llm.ts            # OpenAI-compatible API client with retry + concurrency
│   └── prompts.ts        # All LLM prompts + buildDirectContext
├── embeddings/           # Semantic search (v2.0 phase 2)
│   ├── embed.ts          # Embedding API client
│   ├── vector-store.ts   # Flat-file .embeddings.json
│   ├── indexer.ts        # Auto-index on `remember`
│   └── hybrid-search.ts  # Semantic + keyword + time decay
├── mcp/                  # MCP server (v2.0 phase 1) + config writer (v2.4)
│   ├── server.ts         # stdio transport, handler wiring
│   ├── tools.ts          # remember / recall / search_memories tools
│   ├── resources.ts      # project-context resource
│   └── config-writer.ts  # idempotent merge into .cursor/mcp.json + .windsurf/mcp.json (init --with-mcp)
├── rules/                # Multi-target rules export (v2.4)
│   └── agents-md-writer.ts  # Pure idempotent merge into AGENTS.md (managed-section markers)
├── git/                  # Thin wrappers around the user's `git` binary (v2.4)
│   └── log-reader.ts     # parseGitLog / parseBulkLog (pure) + isGitRepo / getFileHistory / getRecentCommits / isPathTracked (execFile)
├── dashboard/            # Local web UI (v2.1)
│   ├── server.ts         # node:http API (/api/stats, /api/memories, /api/conversations, /api/quality, /api/graph)
│   └── html.ts           # Embedded SPA (Tailwind + D3.js)
├── bundle/               # Portable export/import (v2.3)
│   └── bundle.ts         # Versioned JSON schema v1 + Zod validation
├── store/
│   ├── memory-store.ts   # Read/write Markdown memory files
│   └── state.ts          # Tracks which conversations were processed
├── output/
│   └── terminal.ts       # ANSI colors (respects NO_COLOR), formatting
└── utils/
    ├── author.ts         # Author resolution: CLI > config > git > OS
    └── scheduler.ts      # Cross-platform scheduled-task registration (launchd/crontab/schtasks, v2.6)

bench/
└── cceb/                 # Cursor Conversation Extraction Benchmark (v2.4)
    ├── types.ts          # Pure types: Fixture / ExpectedMemory / Scorecard
    ├── scorer.ts         # Pure scoring (greedy keyword match → P/R/F1) + Markdown render
    ├── loader.ts         # Read + validate fixtures/*.json (hand-rolled, no Zod)
    ├── runner.ts         # Glue: fixture → real extractMemories() → scoreFixture
    ├── run.ts            # CLI entry (`tsx bench/cceb/run.ts [--dry-run] [--filter ...]`)
    ├── fixtures/         # 9 hand-curated annotated conversations (5 types + CJK + 2 noise)
    └── README.md         # Methodology + how to author fixtures

docs/assets/demo/        # Hero GIF source of truth (v2.4) — vhs-rendered, deterministic
├── demo.tape            # 5-frame ≈30s vhs cassette (rendered by `npm run demo:render`)
├── scenario/            # Hand-curated .ai-memory/ store the cassette runs against
│   └── .ai-memory/      # 1 decision + 1 convention + 1 architecture, 2 authors, English-pinned
├── RECORDING.md         # Install matrix (macOS / Linux / Windows / Docker) + pre-commit checklist
└── demo.gif             # Rendered output — committed by maintainer, not CI