Skip to content
/
Phase D Batch 14

Batch 14 — Memory Secondary (Planning, Sessions, Knowledge Capture)

Batch 14 — Memory Secondary (Planning, Sessions, Knowledge Capture)

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
planning-with-files 22,134 claude-plugin no no none no A
claude-code-plan-export 21 claude-plugin no no none no A (archived)
grov 192 npm-package yes (grov) cloud-dashboard none yes (haiku drift-scorer) A
corsair 759 mcp-server no web-dashboard none no A
dev-workspace-dilberry 39 standalone-repo no no parallel-fan-out no A
continuous-claude-parcadei 3,789 standalone-repo no no hierarchical yes (sonnet daemon) A
design-os unknown unknown unknown unknown unknown unknown C
knowledge-flywheel unknown unknown unknown unknown unknown unknown C
second-brain-nicholas 355 skill-pack no obsidian sequential no A
pai-personal-ai-infra 14,434 bash-script-bundle no web-dashboard (31337) hierarchical yes (sonnet+perplexity+gemini+grok+elevenlabs) A

Intra-batch Patterns

This batch reveals a clear spectrum from "minimum viable state persistence" to "maximum ambient intelligence." Planning-with-files, claude-code-plan-export, and dev-workspace-dilberry all implement simple file-based working memory with different scope (project/flat vs branch-scoped) — each adds hooks but no database. Grov and Continuous Claude both cross the database threshold (Supabase+SQLite and PostgreSQL+pgvector respectively) and both extract reasoning from agent sessions, but Grov focuses on team sharing while Continuous Claude focuses on individual compounding. PAI is in a different category entirely — it is the only "life OS" framing in the entire corpus, combining the Algorithm's epistemological framework with 45 skills, a background daemon, and multi-modal output. The batch also reveals how narrow "memory" is as a category: second-brain is PKM, Corsair is integration infrastructure, and knowledge-flywheel/design-os are Reddit rumors without confirmed repos.

Most Interesting Finds

  1. planning-with-files (22K stars): The SHA-256 tamper-attestation system for plan files is a novel security primitive not seen anywhere else — guards against prompt injection attacks through the planning files themselves. The 17-IDE adapter matrix and POSIX portability work represent remarkable ecosystem coverage for a skills-only tool.

  2. PAI (pai-personal-ai-infra): The Sonnet prompt classifier that decides MODE+TIER before every response via a UserPromptSubmit hook is architecturally unique — the model does not decide its own output format, an external classifier does. Combined with the closed 19-item thinking-capability enumeration (phantom names = CRITICAL FAILURE), this creates the most formally constrained response contract in the corpus.

Items Written as Tier C

  1. design-os — Reddit mention only; no confirmed public repository. Concept: "onboarding into vibe-coded monolith." UNCONFIRMED per reddit-mentions-verified.md.

  2. knowledge-flywheel — Reddit mention only; no confirmed public repository. Concept: "session-compounding learnings extractor." The concept is implemented by multiple other frameworks (Continuous Claude's compound-learnings skill, Grov's reasoning extraction) but no standalone product confirmed.

Cross-References Discovered

  • planning-with-files has 5 confirmed community forks (devis, multi-manus-planning, plan-cascade, agentfund-skill, openclaw-github-repo-commander) and is cited by 3 named projects in the wild (ClarityFinance, vv-claude-harness, CCteam-creator)
  • claude-code-plan-export explicitly superseded by native Claude Code planDirectory setting — the only framework in the batch confirmed deprecated by platform feature parity
  • Continuous Claude is conceptually inspired by ccmemory (the memory seed) and superpowers; its YAML handoff concept may be the origin pattern for knowledge-flywheel
  • PAI was created by Daniel Miessler, who also created Fabric (ExtractWisdom pattern used as a PAI skill). The Fabric patterns appear as first-class PAI capabilities.
  • second-brain explicitly implements Andrej Karpathy's LLM Wiki pattern (cited gist) — the only framework in the batch with a clearly named academic/practitioner origin
  • grov's anti-drift detection (monitoring agent actions vs stated goal via Haiku) is conceptually related to what PAI's PromptProcessing.hook.ts does (classify intent before every response)