memory-search

name	memory-search
version	0.1.0
description	Tier 1 semantic memory search across Serena and Forgetful with progressive disclosure and token-budget warnings. The focused search operation split out of the memory router per ADR-063. Use when you say `search memory`, `what do we know about X`, or `recall prior context`. Do NOT use to extract session episodes, update the causal graph, or add citations (use memory or memory-enhancement).
license	MIT
model	claude-sonnet-4-6
metadata	{"adr":"ADR-007, ADR-037, ADR-038, ADR-056, ADR-063","type":"operation","parent":"memory"}

Memory Search

Tier 1 semantic search over memory: the focused search operation extracted from the memory router per ADR-063 (memory-skill decomposition). memory still routes here; a caller that needs only search loads this sub-skill instead of the full memory surface.

Triggers

Use this skill when the user says:

• search memory for semantic search across stored knowledge
• what do we know about X to recall facts, patterns, or constraints
• recall prior context before changing existing code or architecture

Quick Start

# Search memory (Tier 1)
python3 .claude/skills/memory/scripts/search_memory.py "git hooks"

# Lexical-only fallback when Forgetful is unreachable
python3 .claude/skills/memory/scripts/search_memory.py "git hooks" --lexical-only

# JSON output for scripting
python3 .claude/skills/memory/scripts/search_memory.py "git hooks" --format json

The search script is canonical and shared with the memory router; this sub-skill does not reimplement it. It lives at .claude/skills/memory/scripts/search_memory.py.

Quick Reference

Operation	Script	Key Parameters
Search facts/patterns	search_memory.py	query, --lexical-only, --max-results, --format

Routing

search_memory.py "<q>" keyword-ranks Serena memory names (Serena-first, augments with Forgetful when reachable, flags large memories by token estimate) and returns the relevant *-index. Read that index, then follow its links to the atomic file.

Raw fallback when scripting: guess read_memory("<intuitive-name>") (a miss is a cheap "not found", not a list), then the domain *-index, then read_memory("memory-index"). Prefer these name and index lookups over a bare list_memories.

Progressive Disclosure

This skill implements progressive disclosure: list names, read details, deep dive on cross-references. The point is to avoid loading a large memory when a small slice answers the query.

Layer	Tool	Cost	When to Use
Index	search_memory.py	~100-500 tokens	Always start here
Details	read_memory	~500-10K tokens	After index confirms relevance
Deep Dive	Follow cross-references	Variable	For complete understanding

Progressive disclosure prevents loading 9,500 tokens when only 1,200 are relevant. List names first, then read only the entry the query needs.

Graceful Degradation

Search degrades, it does not fail. When Forgetful is unreachable, pass --lexical-only to run Serena-only lexical search. Serena is the canonical store and travels with the repository; Forgetful is supplementary and local. A search that loses Forgetful returns Serena results plus a coverage note, never an error (ADR-007).

Decision Tree

What do you need?
│
├─► Current facts, patterns, or rules?
│   └─► search_memory.py "<query>"
│
├─► Forgetful unreachable?
│   └─► search_memory.py "<query>" --lexical-only
│
└─► Not sure which tier?
    └─► Start here (Tier 1), escalate to the memory router if insufficient

Verification

Operation	Verification
Search completed	Result count > 0 OR logged "no results"
Forgetful degraded	Coverage note present; Serena results returned

Anti-Patterns

Anti-Pattern	Do This Instead
Skipping memory search	Always search before multi-step reasoning
Loading full memory eagerly	List names first, read only the relevant entry
Forgetful hard dependency	Use --lexical-only fallback

Process

Phase 1: Query

Run search_memory.py with the query. Choose --lexical-only when Forgetful is known unreachable.

Phase 2: Validate

Verify results are non-empty and relevant to the query context. A zero-result search is a valid outcome; log it.

Phase 3: Report

Return the relevant index name and the atomic entries it links, with source attribution, to the caller.

Related Skills

Skill	When to Use Instead
memory	Router for episode extraction, causal graph, health, or maintenance
memory-enhancement	Add citations, verify code references, track confidence
curating-memories	Memory maintenance (obsolete, deduplicate, link)
exploring-knowledge-graph	Multi-hop graph traversal beyond Tier 1 search

Troubleshooting

Symptom	Cause	Recovery
Zero results	Query too narrow or memory index empty	Broaden query terms; verify memory index contains relevant entries
Forgetful unreachable	MCP server down or not configured	Use --lexical-only fallback; search continues with Serena only
Slow response	Large memory corpus or network latency	Consider query pagination; check Forgetful server health
Partial results	Token budget hit mid-retrieval	Results include coverage note; caller decides whether to continue

Extension Points

Extension	How to Add
New memory tier	Create sibling sub-skill (e.g., memory-episode); register in memory router
Custom search backend	Implement script matching search_memory.py interface; update skill to invoke
Result post-processing	Chain output through memory-enhancement for citations or confidence scoring

References

• ADR-007: Memory system architecture
• ADR-037: Memory tier boundaries
• ADR-038: Episodic memory structure
• ADR-056: Memory progressive disclosure
• ADR-063: Memory skill decomposition (this extraction)