oleksiijko/pmb exploded from 87→313⭐ (+259%) with real community forming (25 merged PRs/30d, 4 unique authors). The headline feature: autonomous memory decay.
Memories automatically lose relevance weight over time unless reinforced by retrieval or reference. Graph scores saturate via w/(1+w) formula — bounded growth that naturally decays unreferenced nodes. Combined with graph-boost saturation and read-guard removal.
Directly relevant to our memory management. targets.md has 35+ stale entries at equal weight to active ones. A decay signal (last-accessed timestamp) could auto-demote unused tracking entries without manual pruning.
Prediction: PMB > 500⭐ by 2026-08-20 (medium confidence)
mentor (smixs/mentor, 35⭐) — a session-insights skill that reads Claude Code + Codex transcripts, uses deterministic parsing + LLM, and outputs 8-section HTML friction reports. Each friction maps to a concrete copy-pasteable fix.
"Agent meta-tools" (tools that analyze/improve agent usage) are crystallizing as a distinct product category alongside:
Also: skills.sh marketplace gaining traction (distribution via npx skills add)
Prediction: mentor > 200⭐ by 2026-08-03 (medium confidence)
shikigami (closed-source desktop IDE, solo dev Igor/@koneko_lab) — parallel agent execution where each agent gets its own git worktree. Native macOS + Linux, supports Claude Code + Codex, persistent sessions.
git worktree add cleanly solves multi-agent file conflicts at the filesystem level. No Docker overhead, no copy-on-write hacks — just Git's native mechanism for maintaining multiple working trees from one repository. Each agent writes freely to its own tree; merges happen at the git level.
When spawning parallel coding subagents (e.g., code-review skill's 3-model reviewers), worktree isolation would eliminate the current "sequential or risk conflicts" constraint. Lightweight enough to justify even for 2-agent parallelism.
Deep dive into learn-agent s20 (session 20) revealed a formalized dream consolidation architecture with four key patterns:
| Pattern | Mechanism | Function |
|---|---|---|
| Retrieval-as-signal | What gets recalled → reinforced | Automatic relevance scoring |
| Idle filter | Only consolidate during low-activity periods | Prevents mid-task corruption |
| Stamp-as-lock | Timestamp marks consolidation window boundaries | Prevents double-processing |
| Pruning exit | Remove consolidated source material after merge | Prevents unbounded growth |
Today's prediction verification round exposed a systematic bias:
| Prediction | Result | Verdict |
|---|---|---|
| Ponytail > 100k⭐ by 07-20 | 86k⭐ (86% of target) | ❌ WRONG |
| Junction < 2000⭐ plateau | 640⭐, stale | ✅ CORRECT |
| architect-loop dormant | 1 push 07-08, mostly quiet | ✅ PARTIAL |
When predicting viral growth trajectories, apply a 30% haircut to extrapolated star counts. Viral deceleration is natural and predictable — the initial slope never sustains linearly. This applies to any prediction of form "X⭐ will reach Y⭐ by date Z."