Fireworks Skill Memory
https://github.com/yizhiyanhua-ai/fireworks-skill-memory- Category
- Developer Tools
- Rank
- No. 1450Tools index
Previous survey · No. 1437 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- yizhiyanhua-ai
- GitHub
- 93 stars
- Latest release
- v3.0.0
- Date
About
Persistent, per-skill experience memory for Claude Code.
What it does
It turns mistakes and useful discoveries from coding-agent sessions into reusable lessons. Claude Code hooks capture errors, distil lessons after sessions, and insert relevant entries before later skill use. Codex shares the storage model but requires explicit commands to inject lessons, record checkpoints, and process summaries.
Why it's ranked here
The design solves a real repetition problem with scoped, inspectable Markdown rather than opaque model changes. Automatic Claude Code hooks make the loop practical, while shared storage gives Codex partial access. The runtime experience is uneven, though: Codex requires manual workflow steps, and automated distillation depends on a configured model backend.
What's good
Lessons stay scoped to individual skills, limiting irrelevant context. Injection favors frequently matched entries, while capacity management penalizes older, low-hit material. Error capture covers general tool results, not only skill reads. The project also separates storage, transcript parsing, distillation, and runtime adapters, making its main mechanisms understandable and replaceable.
Tradeoffs
Claude Code receives the complete automatic loop, but Codex users must invoke commands and provide structured lesson sections in summaries or session exports. Error detection relies on keyword patterns, so routine but valuable discoveries may never reach automatic distillation. Duplicate detection compares an early text fragment, which can merge distinct lessons sharing similar openings. Optional OpenAI distillation can send prompts to a configured API.
How to use it well
Use it when teams repeatedly run the same Claude Code skills and want proven fixes available during future planning. Keep lessons concrete, narrowly scoped, and phrased with distinctive openings. Codex users should make checkpointing and end-of-task summaries part of their routine. It does not replace general project documentation, broad conversational memory, or a full automatic Codex integration.
Technical notes+
memory_core/store.py stores Markdown entries, ranks injection by [HIT:N], merges approximate duplicates using the first 32 normalized characters, and evicts entries using hits plus an age penalty. memory_core/transcript.py parses Claude JSONL logs and detects skill use and error signals. scripts/pre-skill-inject.py, scripts/inject-skill-knowledge.py, scripts/error-seed-capture.py, and scripts/update-skills-knowledge.py implement Claude lifecycle hooks. cli/skill_memory.py exposes inject, checkpoint, show, and flush; runtimes/codex_runtime.py extracts lesson sections from Markdown, text, JSON, JSONL, or stdin. memory_core/distiller.py supplies Claude CLI, OpenAI-compatible chat-completions, and null backends.
Observed
- License
- MIT License
- Primary language
- Python
- Runtime requirement
- Python 3.9 or newer
- Interfaces
- Claude Code lifecycle hooks and a command-line interface for Codex or manual use
- Install surface
- Shell installer, Codex setup script, and global installation through npx skills
- Platform support
- Claude Code and Codex
- Storage format
- Local Markdown knowledge and checkpoint files, with JSON files used for selected runtime state
Read from README.md, cli/skill_memory.py, memory_core/store.py, memory_core/__init__.py, memory_core/distiller.py, memory_core/transcript.py, runtimes/codex_runtime.py, runtimes/claude_runtime.py, scripts/pre-skill-inject.py, scripts/error-seed-capture.py, scripts/inject-skill-knowledge.py, scripts/update-skills-knowledge.py, LICENSE, SKILL.md.
What it can do
Store skill-specific experiences and learnings
Code interactions and outcomes → Persistent memory records per skill
Retrieve relevant past experiences for current tasks
Current coding task or skill context → Related historical interactions and solutions
Build cumulative knowledge across coding sessions
Multiple coding interactions over time → Accumulated expertise database per skill area
Improve code suggestions based on historical context
Previous successful code patterns and user preferences → Enhanced code recommendations
Track skill development progress over time
Coding attempts and results across sessions → Skill progression metrics and insights
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.