SkillForge

~3000 → ~80 tokens (~50×)

The problem

A procedural memory layer for AI coding agents. When an LLM fixes a bug, the fix-pattern is distilled into a reusable Claude Skill file; AST-based retrieval surfaces it next time the same error class appears; a human gates what enters the library (MCP-served).

The constraint

Context cost. Re-deriving a known fix burns ~3000 tokens of exploration; retrieving the distilled skill costs ~80. Target: ~50× reduction on repeat-class errors.

What the git log says

• $ Episodic memory (mem0-style, vector-backed) remembers what happened; procedural memory remembers how to do things — SkillForge bets coding agents need the second kind.
• $ Indexed by code structure (AST) rather than embedding similarity; AST node-type + error-class is a surprisingly strong retrieval key.
• $ Findings from the Hermes/OpenClaw long-running boxes (tool-loop convergence, memory bleed) feed the design.

What I learnt

Retrieval precision beats recall when a human gates the library. Mode-collapse in eval consensus is the open problem. (WIP — present tense by design.)