MemForge Backlog

Issues to create on GitHub once the repository is public. Each section maps to one GitHub issue.

Testing

1. Add mocked LLM provider tests for consolidation, reflection, and sleep cycle

Labels: enhancement, testing

The integration test suite covers core CRUD paths but all LLM-dependent codepaths are untested:

Summarize consolidation — LLM distillation, entity/fact extraction, relationship creation
Reflection — LLM insight synthesis, contradiction detection, procedural extraction
Meta-reflection — Second-order reflection on accumulated reflections
Sleep cycle Phase 3 — LLM memory revision (augment/correct/merge/compress decisions)
Procedural extraction — LLM condition→action rule extraction from reflections
Semantic/hybrid search — Requires embedding provider

Approach: Create mock LLMProvider and EmbeddingProvider implementations that return deterministic JSON matching expected schemas. Use in tests/llm-paths.test.ts.

Acceptance criteria:

All LLM-dependent paths have at least one happy-path test
Mock providers return valid JSON matching each system prompt's expected schema
Tests run against real PostgreSQL
Revision history verified after sleep cycle test

2. Add end-to-end API tests via HTTP

Labels: enhancement, testing

Current tests call MemoryManager directly. Need HTTP-level tests that exercise the full Express stack: auth middleware, rate limiting, request validation, response format, error handling.

Approach: Start the server on a random port, make HTTP requests with fetch(), assert response bodies and status codes. Test auth failures, invalid input, and rate limiting.

3. Add load/performance tests

Labels: enhancement, testing, performance

No performance benchmarks exist beyond the Redis cache microbenchmark. Need tests for:

Query latency at various warm-tier sizes (1K, 10K, 100K rows)
Consolidation throughput (rows/second)
Sleep cycle duration vs. dataset size
Concurrent query handling

Infrastructure

4. Add GitHub Actions CI/CD pipeline

Labels: enhancement, infrastructure

No CI/CD pipeline exists. PRs can be merged without type-checking or testing.

Required jobs:

type-check — npm run type-check on Node 22
test-integration — with PostgreSQL service container (pgvector/pgvector:pg16)
test-cache — with Redis service container
lint — ESLint

Service containers: pgvector/pgvector:pg16 (includes pg_trgm), redis:7-alpine

Triggers: Push to main, all PRs. Matrix test on Node 20 and 22.

5. Publish to npm as @salishforge/memforge

Labels: enhancement, infrastructure

Package.json is configured with exports, bin entries, and files list. Needs:

npm account setup for @salishforge scope
GitHub Actions publish workflow (on tag push)
Verify npm pack includes only intended files
Test installation in a fresh project

Performance

6. Streaming consolidation for large hot-tier backlogs

Labels: enhancement, performance

consolidate() loads all pending hot-tier rows into memory at once (up to CONSOLIDATION_BATCH_SIZE, default 500). For agents with large backlogs (10K+ events), this causes high memory usage.

Approach: Use cursor-based pagination — process 50 rows at a time, commit each batch independently. Requires breaking the single-transaction model into per-batch transactions with idempotent re-runs.

Challenges:

Transaction boundaries: currently entire consolidation is one transaction
Error recovery: partial progress must be preserved on crash
LLM calls remain the real bottleneck regardless of streaming

7. Connection pool tuning and health checks

Labels: enhancement, performance

The PostgreSQL connection pool uses default pg.Pool settings. For production workloads:

Pool size should auto-scale based on concurrent requests
Idle connection cleanup should be configured
Connection health checks should detect stale connections
Pool exhaustion should return clear errors, not hang