Skip to content
/
Phase B Batch 25

Batch 25 Notes — Yegge Beads ecosystem + FlowCoder + persistent agent-OS

Batch 25 Notes — Yegge Beads ecosystem + FlowCoder + persistent agent-OS

Roster

Slug Stars Lang Dist Hooks Skills MCP Tools Status
gas-town 15,581 Go cli-tool 0 4 0 Complete
beads-ui 621 TypeScript npm-package 0 0 0 Complete
superbeads-wiggum 31 Bash bash-script-bundle 0 6 commands 0 Complete
flowcoder 26 Python standalone-repo 0 5 JSON cmds 0 Complete
onebrain 10 Markdown claude-plugin 1 34 0 Complete
memsearch-cc 1,841 Python claude-plugin 4 1 0 Complete
mcp-shrimp-task-manager 2,103 TypeScript mcp-server 0 0 16 Complete
agentmemory 18,016 TypeScript npm-package 12 8 53 Complete
mempalace 52,851 Python python-package 2 1 30 Complete
ai-afterimage 22 Python python-package 2 0 0 Complete

Canonical Source Resolution

  • beads-ui: Reddit ref — found via gh search repos beads-ui. mantoni/beads-ui (621 stars) matched the "beads JSONL kanban UI" description.
  • superbeads-wiggum: Reddit ref — found via gh search repos superbeads. EliaAlberti/superbeads-universal-framework (31 stars) was the only result with explicit "Wiggum Flavour" branding.
  • memsearch-cc: Reddit ref — found via gh search repos memsearch. zilliztech/memsearch (1,841 stars) — Zilliz is the Milvus company, confirming this is the Milvus-backed memory search tool referenced in awesome-claude-code.
  • All other repos found directly by GitHub URL or org/repo slug.

Intra-Batch Analysis

Star Range

This batch spans the widest star range in the Phase B corpus: from 10 (onebrain) to 52,851 (mempalace). The top three (mempalace 52k, agentmemory 18k, gas-town 15k) are among the highest-starred frameworks in the entire corpus.

The Beads Sub-Ecosystem

gas-town, beads-ui, and superbeads-wiggum form a coherent Beads sub-ecosystem:

  • beads (bd/bv CLI by mantoni, not in this batch) is the JSONL issue tracker
  • beads-ui (mantoni, same author) is the web visualization layer
  • gas-town (Yegge) is the multi-agent coordinator that uses Beads as its issue tracker
  • superbeads-wiggum (EliaAlberti) is a community fork that adds CLAUDE.md integration and domain packs to the Beads CLI

Yegge's Beads philosophy ("everything is a Molecule, root sources only, no nudging") is the conceptual foundation, but each tool takes it differently: beads-ui is purely visual, gas-town uses it for agent task management, superbeads-wiggum adds slash-command workflow scaffolding.

Persistent Memory Systems (5 of 10)

Five frameworks in this batch are memory systems: agentmemory, mempalace, memsearch-cc, onebrain, ai-afterimage. They form a spectrum:

Dimension agentmemory mempalace memsearch-cc onebrain ai-afterimage
Storage SQLite/KV (iii-engine) ChromaDB + SQLite Markdown + Milvus Obsidian files SQLite/PostgreSQL
Verbatim vs compressed LLM-compressed Verbatim Raw markdown Verbatim Verbatim (code only)
Hooks 12 (all events) 2 (Stop+PreCompact) 4 (all session lifecycle) 1 (SessionStart) 2 (Pre/PostToolUse)
Scope All tool interactions Code + conversations All tool interactions User ideas/tasks Code files only
Knowledge graph Yes (LLM-extracted) Yes (SQLite temporal) No No No
API key required Optional (LLM features) No (raw search) Optional No No

Deny-Then-Allow Hook Pattern

Two frameworks use this pattern: memsearch-cc and ai-afterimage. Both deny the first Write/Edit attempt to force the agent to read injected context. The pattern appears independently invented — memsearch-cc uses a shell script + Python approach while ai-afterimage uses a single Python file with hash-based state tracking.

Dynamic Instructions Pattern

mempalace's skill calls mempalace instructions <command> to fetch live instructions from the CLI at runtime. This is the only framework in the corpus to do this — prompt updates ship with pip upgrade not plugin reinstalls.

Benchmark Competition

agentmemory (95.2% R@5 LongMemEval-S) and mempalace (96.6% R@5 raw, 98.4% hybrid) publish competing benchmarks on the same dataset. mempalace achieves higher precision without any LLM; agentmemory closes the gap with LLM compression enabled. mempalace explicitly refuses to include side-by-side competitor comparisons, while agentmemory publishes a competitor comparison table.

Single-File Architectures

Two frameworks in this batch have notable single-file designs:

  • ai-afterimage: entire hook logic in hooks/afterimage_hook.py (~700 lines)
  • flowcoder: entire pipeline in a Tkinter Python app with a single JSON command format

FlowCoder Uniqueness

flowcoder is the only framework in Phase B that uses a visual desktop GUI (Tkinter) for workflow authoring. It auto-commits after every block execution, creating an audit trail of agent work. No other framework in the corpus auto-commits per-step.

OneBrain Scale

onebrain has 34 skills — the largest skill count in the Phase B corpus. Skills are directories (not flat .md files), each with a SKILL.md. The braindump skill is noteworthy: it classifies free-form user input into 6 categories (Task/Idea/Note/Project/Question/Feeling) and creates inbox files immediately without confirmation.

Cross-Batch Connections

  • agentmemory competes directly with mempalace on LongMemEval benchmarks
  • gas-town uses Beads (batch 25) + beads-ui (batch 25) as its data layer
  • memsearch-cc uses Milvus — same vector database company (Zilliz) that hosts the memsearch repo
  • mcp-shrimp-task-manager is closest to taskmaster-ai (batch 1) but adds process_thought enforcement
  • ai-afterimage parallels memsearch-cc in deny-then-allow pattern but with code-only scope and churn tracking

Tier C Stubs

None required. All 10 frameworks were findable: 7 by direct URL, 3 via gh search repos with at most 2 attempts each.