RadioHeader

Cross-project memory for coding agents — stop re-solving bugs you already fixed in another project.

You: App shows white screen for 10+ seconds on launch

Claude: RadioHeader has experience from ProjectA:
        "Task {} in SwiftUI .onAppear inherits Main Actor —
        iCloud I/O blocks main thread, causing 10s+ white screen.
        Fix: use Task.detached(priority:)"

        Verifying this applies here... ✓ Same pattern. Applying fix.

One install. Works on Claude Code, Codex CLI, and any MCP-compatible agent. Zero config.

中文文档 · How It Works · MCP Server · Lessons Learned

What it does

Coding agents are great at analysis but don't accumulate experience. Even with built-in memory, project memories are completely isolated: what Project A learned, Project B has no idea about.

RadioHeader solves this with three mechanisms:

The result: the second time any agent in any project hits a similar problem, the answer is already there. Minutes to seconds.

What RadioHeader adds to coding agents

How memory flows

Conversation (bug fix, feature, debugging session)
     │
     ▼
 ┌─ Echo (automatic, via hooks) ─────────────────────┐
 │  Claude: PostToolUse Write|Edit → immediate        │
 │  Codex:  UserPromptSubmit snapshot + Stop diff      │
 │  Result: experience written to project memory/      │
 └───────────────────────────┬───────────────────────┘
                             ▼
 ┌─ Three-Layer Memory ──────────────────────────────┐
 │                                                    │
 │  Project memory/ ─── is this cross-project? ──→    │
 │                          yes ↓                     │
 │  RadioHeader topics/ ──── [source:ProjectName] ──→ │
 │                          distill ↓                 │
 │  RadioHeader shortwave/ ── project-agnostic ──→    │
 │                          searchable everywhere     │
 └───────────────────────────┬───────────────────────┘
                             ▼
 ┌─ Search → Apply → Trace ──────────────────────────┐
 │  Next time any agent hits a similar problem:       │
 │  1. Search: Grep with symptom keywords             │
 │  2. Apply:  Cite the matching entry in the answer  │
 │  3. Trace:  Follow [source:] to the origin project │
 └──────────────────────────────────────────────────┘

RadioHeader (~/.claude/radioheader/)
├── shortwave/            ← refined, project-agnostic knowledge
├── topics/               ← detailed experience with [source:] tags
├── project-registry.json ← project cards (domains, problems, pain points)
├── context-digest.md     ← attention-compressed awareness (auto-generated)
└── INDEX.md              ← master index

Memory syncs accumulate → every 5 syncs, consolidate runs automatically
                                    ↓
                  Analyzes search logs + project activity + user profile
                                    ↓
                  Generates context-digest.md (compressed awareness)
                                    ↓
                  Next session: Agent starts knowing who it's helping

📡 Shortwave (local)     — your own experience (highest priority)
🌐 Community (shared)    — entries from other users, with quality scores
📂 Topics (detailed)     — raw topic entries with [source:] tags

pip install "mcp[cli]"
radioheader mcp-server          # starts stdio server

Runtime support

RadioHeader ships a single data layer with runtime adapters that plug into whichever coding agent you're using:

The Codex adapter uses a snapshot + diff loop: UserPromptSubmit fingerprints memory/topics/shortwave before each turn; Stop diffs against that fingerprint and uses decision: block with a continuation prompt if any step in the Echo chain was skipped.

Python hooks target Python 3.7+ (from __future__ import annotations). Snapshot walker caps at 10k files with 24h TTL cleanup.

Setup

git clone https://github.com/ZaptainZ/radioheader.git
cd radioheader
./install.sh --runtime both

That's it. Start Claude Code or Codex in any project — hooks fire, rules are loaded, experience is searchable.

./install.sh --runtime claude   # only Claude Code
./install.sh --runtime codex    # only Codex CLI
./install.sh --runtime both     # default

Optionally, run radioheader init inside a project for per-project scaffolding (Echo rules, log directory, doc templates).

Use

# Search by symptoms, not solutions
radioheader search "white screen|slow launch|startup"

# Initialize a project
radioheader init --name "MyAPI" --stack "Python/FastAPI"

# Learn from an article
radioheader learn https://example.com/article

# Check health
radioheader doctor

Full CLI:

RadioMind integration

RadioMind is a bionic memory core for AI agents — it distills scattered conversations into deep habits through three-body debate and dream refinement, then serves them back when they matter. RadioHeader captures and enforces; RadioMind refines and deepens.

When RadioMind is installed, RadioHeader automatically upgrades itself — no config needed:

RadioMind reads RadioHeader's topics/ and shortwave/ files directly. RadioHeader doesn't need to know RadioMind exists — it just detects radiomind on PATH and delegates search and consolidate. If RadioMind fails or is absent, the native path runs seamlessly.

RadioHeader (capture + rules)              RadioMind (refine + enhance)
─────────────────────────                  ─────────────────────────
Echo hooks → write topics/shortwave/ ────→ adapter reads, indexes, refines
radioheader search ─── radiomind? yes ──→ radiomind rh-search
                   └── no ──→ fts-search.py (native)
radioheader consolidate ── radiomind? ──→ radiomind rh-consolidate
                        └── no ──→ attn-consolidate.py (native)

Community library

radioheader-community is an opt-in shared library of shortwave entries contributed by the RadioHeader community. Your local experience stays local; entries you choose to publish pass three gates before reaching the shared pool.

radioheader community on          # enable
radioheader sync                  # pull shared library + push your votes
radioheader publish <file>        # publish (quality ≥6/8 + privacy + dedup)

Quality governance follows a Stigmergy model — like ant pheromone trails: good entries get reinforced through usage, bad entries decay naturally, no human curation needed.

When enabled, search results blend three sources with clear priority:

📡 Shortwave (local)     — your own experience (highest priority)
🌐 Community (shared)    — entries from other users, with quality scores
📂 Topics (detailed)     — raw topic entries with [source:] tags

Lessons learned

Built through real usage across 13 projects. Four lessons that shaped everything:

"Searched but didn't use" — the #1 failure mode. Claude would search, find results, and ignore them. The fix: three mandatory steps (Search → Apply → Trace) with an explicit prohibition on finding but not using.

Symptom keywords > solution keywords. Developers search "white screen" and "slow launch", not "Task.detached". Every entry must preserve the words someone would actually search for.

Instructions beat knowledge. Writing "experience is stored here" doesn't drive behavior. Writing "you MUST search here first" does. CLAUDE.md / AGENTS.md content must be imperative behavioral rules, not reference documentation.

Attention belongs in compression, not retrieval. Attention-weighted search ranking added zero value — BM25 already gets the right results. The real value is in memory consolidation: compressing the user's project landscape into a digest that shapes how the Agent thinks.

See docs/lessons-learned.md for the full list.

Radio ecosystem

RadioHeader (rules & experience) → RadioMind (memory & habits) → RadioHand (actions & channels)
         head                              brain                          hands

Docs

License

MIT