Semantic search over global agent memory. Use to retrieve previously learned patterns, decisions, gotchas, and workarounds. Prevents stale-context errors across long sessions and multi-agent pipelines.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Semantic search over global agent memory. Use to retrieve previously learned patterns, decisions, gotchas, and workarounds. Prevents stale-context errors across long sessions and multi-agent pipelines.
["Run memory-search at the start of any non-trivial task","Use focused queries, not broad ones","Treat results as pointers to canonical sources, not ground truth","Pair high-similarity results with explicit file reads for critical decisions"]
error_handling
graceful
streaming
supported
source
builtin
trust_score
100
provenance_sha
b17fbd9a4c17f016
Memory Search
Memory Search is the live recall layer for all agents in the ecosystem. It performs semantic search over the global memory corpus (learnings, decisions, issues, gotchas, patterns) using vector similarity. Every agent must use this skill to avoid working from stale or contradictory context — especially in long sessions where in-conversation memory has been compressed or reset.
- Semantic similarity search over all memory files (learnings.md, decisions.md, issues.md, gotchas.json, patterns.json)
- Returns up to 10 results ranked by similarity score
- Output includes: source file, similarity percentage, and 200-character content preview
- Supports natural language queries — no keyword matching required
- Cross-agent knowledge retrieval: surfaces knowledge from previous agents in the same or prior sessions
Step 1: Formulate a Focused Query
Identify the specific knowledge gap before invoking. Good queries target a concrete decision, pattern, or error — not a broad topic.
Good queries:
"JWT refresh token rotation pattern"
"Windows path normalization in hooks"
"BM25 index rebuild performance"
"agent spawning without task_id error"
Bad queries (too broad):
"authentication" — too generic, returns noise
"everything about routing" — unfocused
"what have we done before" — meaningless for semantic search
Found N results for: "<query>"
[<source-file>] Similarity: XX.X%
<200-character content preview>...
[<source-file>] Similarity: YY.Y%
<200-character content preview>...
Example real output:
Found 3 results for: "JWT refresh token rotation pattern"
[learnings.md] Similarity: 87.3%
JWT refresh tokens must be stored in httpOnly cookies to prevent XSS. See ADR-045 for detailed rationale and trade-offs vs localStorage approach...
[decisions.md] Similarity: 74.1%
ADR-045: JWT over sessions chosen for stateless architecture. Refresh token rotation every 15 minutes. Invalidate on logout via token blocklist...
[issues.md] Similarity: 61.2%
Known issue: JWT expiry not propagated to frontend — workaround in auth.middleware.ts line 47 until ADR-045 Phase 2 is implemented...
Step 4: Interpret Similarity Scores
Similarity
Interpretation
Action
>80%
Highly relevant — directly matches the query
Read the full source file for complete context
60–80%
Likely relevant — related topic or adjacent pattern
Skim and use if applicable
40–60%
Possibly relevant — tangentially related
Use judgment, may be noise
<40%
Probably not relevant
Discard unless nothing better exists
Step 5: Follow Up on High-Similarity Results
For results with similarity >60%, read the full source file to get untruncated content:
Check memory files directly if the topic is recent and may not yet be indexed
Zero results means the knowledge is either not yet documented or the query was too specific.
Step 7: Apply or Document Findings
If result confirms known approach: proceed with confidence; cite the source in your reasoning.
If result reveals a better approach: update your plan to align with the documented pattern.
If result reveals a known issue/workaround: apply the workaround immediately rather than discovering it yourself.
If nothing useful found: continue with your best judgment; document your decision afterward to Memory Protocol.
</execution_process>
<best_practices>
Run early, not late: Memory search at the START of a task prevents wasted effort. Running it after you have already made decisions is too late to benefit.
One query per knowledge gap: Don't batch all your unknowns into one query. Run targeted queries for each specific gap.
Use the actual error message as the query: If you hit an error, search for the error message text verbatim — it often finds a documented workaround.
Results are pointers, not canonical sources: Always follow up by reading the full source file when similarity is >70%. The 200-char preview is not the complete record.
After a session gap or context compression: Always run memory-search to re-establish current state before proceeding. Context may have been lost.
Cross-agent knowledge sync: Before duplicating an analysis another agent might have done, search memory. The result may already be documented.
</best_practices>
At the start of a long task or after a context compression event:
# What are the current known issues in this area?
node .claude/lib/memory/memory-search.cjs "hook execution errors"# What patterns have been established for this domain?
node .claude/lib/memory/memory-search.cjs "routing table keyword patterns"# Any decisions I should be aware of?
node .claude/lib/memory/memory-search.cjs "agent model selection ADR"
</usage_example>
<usage_example title="Debugging — check for known workarounds">
When stuck on an error:
# Search for the exact error message
node .claude/lib/memory/memory-search.cjs "Unable to locate module.exports insertion point"# Or describe the symptom
node .claude/lib/memory/memory-search.cjs "routing table update fails silently"
If a result appears with >70% similarity, the error has been seen before and may have a documented fix.
</usage_example>
<usage_example title="Before making an architectural decision">
Before choosing between two approaches:
# What was decided previously about this tradeoff?
node .claude/lib/memory/memory-search.cjs "ESM vs CJS module format decision"# Are there ADRs for this pattern?
node .claude/lib/memory/memory-search.cjs "async hook pattern ADR"
If a relevant ADR is found, follow it rather than re-litigating the decision.
</usage_example>
When you need to know what another agent discovered:
# What did the reflection agent find recently?
node .claude/lib/memory/memory-search.cjs "reflection agent findings skills gap"# What did the security audit reveal?
node .claude/lib/memory/memory-search.cjs "security audit prompt injection findings"
</usage_example>
Iron Laws
ALWAYS run memory-search before assuming no prior context exists. Never claim "there's no documented pattern for this" without first running a targeted search.
NEVER use memory-search results as canonical ground truth. They are summaries and previews. Always read the full source for decisions with material consequences.
NEVER skip memory-search because "I already know about this." Stale in-session assumptions are the primary cause of inconsistent multi-agent behavior.
ALWAYS run multiple targeted queries, not one broad one. A single broad query misses specific matches. Two or three focused queries are more effective.
NEVER run memory-search with a query longer than ~10 words. Over-specified queries degrade similarity matching. Extract the key concept.
Anti-Patterns
Anti-Pattern
Why It Fails
Correct Approach
"everything about auth" broad query
Too generic — returns unrelated noise
"JWT refresh token rotation"
Skipping search because "context is fresh"
In-session context is NOT the same as memory
Always search at task start
Treating 200-char preview as full record
Preview truncates at 200 chars — critical detail may be cut
Follow up with Read on the source file
Running search AFTER making the decision
Memory lookup must happen before your design choice
Search first, decide second
Single query for multiple unknowns
One query can't cover multiple semantic dimensions
