.agents/memory/MEMORY.md

Current Context

Implementing Phase B: Shadow Extraction for the Signet memory pipeline. The focus is building the extraction and decision intelligence layer in shadow mode to validate quality/reliability without destructive consequences.

Active Projects

Phase B: Shadow Extraction Implementation

• Location: packages/daemon/src/
• Status: Planning complete, awaiting implementation
• Dependencies: Phase A infrastructure (schema migrations, DB accessor, feature flags) now in c1e43b6
• Next Steps:
  1. Create src/extract/ module with extractFactsAndEntities(input): ExtractionResult
  2. Build src/decision/ module for shadow candidate retrieval
  3. Implement src/worker.ts with job queue processing (reads from memory_jobs, writes to memory_history)
  4. Add contract validation and warning persistence
• Key Constraints:
  • Extract from daemon.ts recall logic and hooks.ts extraction
  • Keep semantics non-mutating
  • All logs go to memory_history table only
  • New files needed (daemon.ts is 4511 LOC)

Recent Work

Phase A: Infrastructure Hardening (completed in c1e43b6)

• Schema migrations 001+002 for job queue/history tables
• Singleton DB accessor with WAL mode and read pool
• Transaction boundaries keeping provider calls outside write locks
• Content-hash dedup implementation
• Feature flags with kill switches (PipelineV2Config)
• Entity graph schema with proper entity types
• 102 tests passing

Team Structure Established

• Parallel agent delegation used successfully
• Schema-agent, db-accessor-agent, config-test-agent working in sequence
• Integration pass handling remaining wiring

Technical Notes

Database Schema:

• memory_jobs table: stores job queue entries
• memory_history table: stores shadow extraction logs/proposals
• Feature flags: PIPELINE_FLAGS in memory-config.ts

Key Interfaces:

• ExtractionResult: structured facts + entities output
• DecisionEngine: retrieve top-K candidates, confidence scoring
• SignetLifecycle connectors pattern maintained

Tooling:

• Database access via singleton accessor
• Transaction wrappers for write operations
• Content hashing for deduplication
• Vector search integration for candidate retrieval

Code Standards:

• Self-explanatory code with "why" comments
• Max 700 LOC per file, max 3 indentation levels
• Test in browser (not lazy)

Rules & Warnings

• Never delegate UI work to subagents (handle personally as Opus)
• Prefer Sonnet/Haiku delegation when operating as Opus
• Keep journal entries in memory database regularly
• Signet agent profile at ~/.agents/
• Phase B work must be in new files (daemon.ts too large at 4511 LOC)
• Test in browser, don't be lazy
• Non-mutating semantics for shadow extraction phase