refactor(conversation-memory): V3 redesign - 3-hook architecture

[baleen37]

Summary

Complete redesign of conversation-memory plugin from complex observer daemon architecture to simplified 3-hook architecture.

Key Changes

Architecture Simplification:

• Removed observer daemon that polled DB every 1s
• Eliminated dual storage (exchanges + observations)
• Simplified observations schema from 10+ fields to just title + content
• Reduced token injection from ~900 tokens to budgeted 500 tokens

New Components:

• db.v3.ts - Clean slate database schema (pending_events + observations)
• compress.ts - Rule-based tool compression (no LLM)
• observations.v3.ts - Simplified CRUD operations
• batch-extract-prompt.ts - LLM batch extraction for Stop hook
• search.v3.ts - Observation-only search
• post-tool-use.ts - Hook to store compressed events
• stop.ts - Hook for batch LLM extraction at session end
• session-start.ts - Hook for token-budgeted injection
• read.ts - Raw conversation access from JSONL

Legacy Files Deleted:
49 files (~16,000 lines) including:

• parser.ts, sync.ts, tool-compress.ts
• observer.ts, observation-prompt.ts
• round-robin-provider.ts
• All associated tests

3 New Hooks:

1. PostToolUse - Stores compressed tool events (async, no LLM)
2. Stop - Batch LLM extraction from pending_events (10-20 events per batch)
3. SessionStart - Token-budgeted injection of recent observations

MCP Server Simplified to 3 Tools:

1. search - Single query, returns compact observations (Layer 1)
2. get_observations - Full details by ID array (Layer 2)
3. read - Raw conversation from JSONL (Layer 3)

Test Plan

• All 603+ tests passing
• 18 integration tests covering end-to-end workflows
• All components have unit tests
• CI tests passing

Breaking Changes

Config file format changed:
Old format:

{
  "provider": "gemini",
  "gemini": {
    "apiKeys": ["key1", "key2"],
    "model": "gemini-2.0-flash"
  }
}

New format:

{
  "apiKey": "your-gemini-api-key",
  "model": "gemini-2.0-flash"
}

Database migration:

• Old database file deleted on first run (no migration)
• Clean slate with new schema

🤖 Generated with Claude Code

Summary by CodeRabbit

• Tests
  • Expanded integration test coverage for V3 architecture, including end-to-end workflows, session management, search functionality, and error handling scenarios.

[coderabbitai]

Caution

Review failed

The pull request is closed.

📝 Walkthrough

Walkthrough

A comprehensive integration test suite for the V3 architecture is added, exercising end-to-end workflows, session injections, MCP tool integration, batch processing, threshold-based extraction, filtering, and error handling scenarios with mocked LLM providers and in-memory SQLite.

Changes

Cohort / File(s)

Summary

V3 Integration Test Suite plugins/conversation-memory/src/integration.v3.test.ts

New comprehensive test file (778 lines) covering end-to-end workflows, session start hook injection with token budgeting, MCP tool search and retrieval, batch processing, threshold-based extraction, cross-project isolation, and error handling. Includes setup/teardown with in-memory SQLite database, mocked LLM provider and embeddings module, and extensive verification of observations, session injections, and search results.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

Suggested labels

enhancement

Poem

🐰 A burrow of tests now guards the v3 way,
Mocking LLMs and checking day by day,
Sessions injected, observations flow free,
Cross-project isolation—secure as can be! 🔒✨

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch chore-claude-mem3-pr

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}