memory-hook-pre-spawn-pack

name	memory-hook-pre-spawn-pack
description	Hook B — Pre-Spawn Memory Pack build. Runs before each worker spawn (executive Phase 4 prep). Reads the second brain for project-specific memories and produces a Memory Pack markdown block that worker-spawner injects into the worker's generated CLAUDE.md. Workers consume this as static markdown — they never call mem0 themselves. Read-only. Invokes the memory-reader skill.

Hook B — Pre-Spawn Memory Pack Build

You are the executive agent. A worker is about to be spawned for a specific work item. Your job: build a Memory Pack that gets injected into the worker's generated CLAUDE.md so the worker starts with relevant prior context baked in.

The worker will NEVER call mem0 itself. Workers consume static markdown. This is the only way prior memory reaches a worker.

Read first

1. Read .claude/skills/memory-reader/references/mem0-limitations.md.
2. Then invoke the memory-reader skill — your full guide for retrieval lives there.

Context for this consultation

{{CONTEXT_JSON}}

The context block contains:

• workItem — full details: id, title, description, priority, bundle_slug, app_id, output_path if resuming
• currentStep — if this is step execution, the step number, title, and description
• executionPattern — plan-then-execute / harness / etc.
• vendor — which worker vendor (claude / codex / kimi) is about to run
• retryAttempt — if this is a retry, the attempt number and prior strategy

What to retrieve

For the worker to do its best work, surface:

1. Project-specific memories — search filtered to app_id: <bundle_slug>. What's been tried, what's currently the state of this codebase, prior step outcomes.
2. Capability memories — what does the executive know about the capabilities this work item exercises (vendor: claude vs codex, harness X, skill Y)?
3. Principles that apply — any metadata.type: "principle" memories that constrain what the worker can do.
4. Recent failures on this bundle — if retryAttempt > 0, prior failure signals are critical.

Iterative search (3–8 queries from different angles). Stop when results stop changing.

Output shape

Use Shape B from memory-reader/SKILL.md — the Memory Pack format. The TypeScript wrapping this hook will splice your output verbatim into the worker's CLAUDE.md.

## Memory Pack

The executive agent has consulted prior runs and retros for this work. The most relevant prior knowledge:

### [<type>] <60-char title>
**ID:** `<mem_id[:8]>` · **Score:** `<0.42>` · **Source:** `<metadata.source>`
<full text from mem0, verbatim>

### [<type>] <next>
...

Budget — strict

• ≤2K tokens total for the entire Memory Pack section
• Drop any memory with score < V3_MEM0_CONFIDENCE_FLOOR (default 0.7)
• Sort surviving memories by score descending; truncate from the bottom if the budget overflows
• Max 10 entries

If the pack is empty

If no memories meet the confidence floor (genuinely new project, or memory layer is sparse), output:

## Memory Pack

_No prior memories surfaced for this work item. (Searched: <list of queries>.)_

Empty packs are legitimate and the worker handles them fine. Do not fabricate memories to fill space.

Hard rules

• Read-only. No writes.
• Verbatim text from mem0 (do not re-paraphrase — mem0 already paraphrased on extraction; double-paraphrasing destroys discriminators).
• Cite the source path from metadata.source so a worker can read the original markdown if needed.
• Cap total output at 2K tokens. The TS wrapping enforces this as a hard limit and truncates with ...[truncated] if you overflow.