Fireworks Skill Memory

fireworks-skill-memory

Persistent experience memory for Claude Code and Codex skills. Shared memory core, runtime-specific adapters, skill-scoped lessons.

What It Does

Every coding-agent session starts from zero. The same mistakes repeat — wrong API parameters, broken sequences, proxy pitfalls — because the runtime has no durable skill memory between sessions.

fireworks-skill-memory solves this by:

1. Injecting past experience when a skill is invoked
2. Distilling new lessons into skill-scoped knowledge files
3. Keeping runtime-specific adapters thin so Claude hooks and Codex explicit flows share one memory core

Installation

Quick Install (Recommended)

In Claude Code, say:

> "Help me install fireworks-skill-memory from https://github.com/yizhiyanhua-ai/fireworks-skill-memory"

Or run the one-command installer:

curl -fsSL https://raw.githubusercontent.com/yizhiyanhua-ai/fireworks-skill-memory/main/install.sh | bash

npx skills Install

npx skills add yizhiyanhua-ai/fireworks-skill-memory -g

After installing via npx skills, run the installer to set up hooks:

curl -fsSL https://raw.githubusercontent.com/yizhiyanhua-ai/fireworks-skill-memory/main/install.sh | bash

Codex Setup

Run:

./install-codex.sh

Then use:

python3 cli/skill_memory.py inject --skill 
python3 cli/skill_memory.py checkpoint --skill  --note "..."
python3 cli/skill_memory.py flush --skill  --summary-file ./session-summary.md

How It Works

Claude Code installs 4 hooks that run automatically:

| Hook | Trigger | Script | Purpose | |------|---------|--------|---------| | PreToolUse | Before Skill call | pre-skill-inject.py | Inject full KNOWLEDGE.md before skill executes | | PostToolUse | After Read SKILL.md | inject-skill-knowledge.py | Inject top-N entries by relevance + capture error seeds | | PostToolUse | After any tool call | error-seed-capture.py | Capture error signals to session-scoped file | | Stop | Session end (async) | update-skills-knowledge.py | Distill new lessons via Haiku, update KNOWLEDGE.md |

Codex uses explicit runtime commands instead:

| Command | Purpose | |---------|---------| | inject | Load top-ranked lessons for a skill before a task | | checkpoint | Save raw notes into the skill directory | | flush | Distill explicit lesson sections from summary/session inputs into KNOWLEDGE.md |

Data Flow

Skill invoked → PreToolUse injects experience → Claude executes with context
                                                    ↓
Session ends → Stop hook reads transcript → Haiku distills 1-3 lessons
                                                    ↓
                              KNOWLEDGE.md updated → Ready for next session

Codex task starts → inject loads top lessons → Codex executes with context
                                                ↓
Checkpoint/summary captured → flush distills explicit lessons
                                                ↓
                      KNOWLEDGE.md updated → Ready for next session

Knowledge Storage

/skills//KNOWLEDGE.md   ← Distilled per-skill experience
/skills//CHECKPOINTS.md ← Raw runtime notes
/global/KNOWLEDGE.md                ← Global cross-skill principles

Claude legacy runtime also uses:
~/.claude/skills//KNOWLEDGE.md
~/.claude/skill-usage-stats.json
~/.claude/skill-memory.log

Each entry is tagged with [YYYY-MM] timestamp and [HIT:N] usage counter. Low-frequency, old entries are evicted first.

Configuration (Optional)

All settings are optional, configured via environment variables:

| Variable | Default | Description | |----------|---------|-------------| | SKILLS_KNOWLEDGE_MODEL | claude-haiku-4-5 | Default model for the Claude CLI distiller backend | | SKILLS_DISTILLER_BACKEND | claude-cli | Distiller backend selector: claude-cli, openai, or null | | SKILLS_DISTILLER_DEBUG | unset | Enable backend debug logging to the default distiller log file | | SKILLS_DISTILLER_LOG | unset | Explicit path for distiller debug logs | | SKILL_MAX | 100 | Max entries per skill | | GLOBAL_MAX | 100 | Max global entries | | MIN_TOOL_CALLS | 5 | Skip sessions with fewer calls (likely summaries) | | SKILLS_INJECT_TOP | 20 | Max entries injected per active invocation |

Requirements

• Python 3.9+
• Claude Code CLI for the automatic Claude runtime
• Claude Haiku access for the default Claude CLI distiller backend
• CODEX_HOME for the Codex runtime setup helper
• OPENAI_API_KEY / OPENAI_BASE_URL / OPENAI_MODEL when using the OpenAI distiller backend
• SKILLS_DISTILLER_DEBUG or SKILLS_DISTILLER_LOG when debugging backend failures

More Information

• [Full Documentation](README.md)
• [中文文档](README.zh-CN.md)
• Report Bug

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yizhiyanhua-ai
• Source: yizhiyanhua-ai/fireworks-skill-memory
• License: MIT

Install and usage instructions live in the source repository linked above.