| name | context-manager |
| description | Design context-management systems for AI-powered products — RAG, vector stores, knowledge graphs, episodic/semantic memory, multi-agent context handoff. Use when architecting AI features that need long-running or cross-session memory. Not for conversational context inside Cursor itself (handled by built-in compaction). |
| disable-model-invocation | true |
Context Manager
Cursor has built-in context compaction (preCompact hook) and the explore subagent for codebase context. This skill is about building context systems for your own AI products, not about managing Cursor's own context window.
Design context systems that are relevant, fresh, auditable, and cost-aware.
Memory taxonomy
| Type | Purpose | Typical store |
|---|
| Working memory | Current turn / session | LLM context window |
| Episodic | Past interactions with this user | Vector + metadata store |
| Semantic | General facts / domain knowledge | Vector + keyword + KG |
| Procedural | How-to patterns / tools | Structured definitions |
Design axes
- Relevance — retrieval precision > recall for LLMs; rank ruthlessly.
- Freshness — stale context poisons answers; track write time and TTL.
- Tenancy — multi-tenant isolation at retrieval, not just storage.
- Compliance — PII classification, audit of which context produced which answer.
- Cost — token cost of every retrieved chunk; measure the marginal value.
Response workflow
- Requirements — what does the AI feature need to remember, about whom, for how long?
- Architecture — pick stores (vector, KG, SQL, object), integration surface, caching layer.
- Dynamic assembly — given a query, how is the context pack built? Rules + retrieval + reranker.
- Optimise — compression / summarisation, dedup, hierarchical retrieval.
- Integrate — clean API surface between retrieval and generation.
- Measure — relevance @ k, answer quality with/without retrieved chunks, cost per query.
- Iterate — A/B different retrievers, chunk sizes, rerankers.
- Scale — sharding, read replicas, async indexing.
- Document — data flow, lineage, retention, compliance.
- Evolve — retire stale sources; update as the product learns.
Enterprise considerations
- Multi-tenant isolation at the query level (tenant ID in every filter).
- Source integrations: Confluence, SharePoint, Notion, Drive; respect source-level ACLs.
- Lifecycle: archival, deletion on user request (GDPR / right to erasure).
- Audit trail: every answer links back to the chunks that produced it.
Anti-patterns
- "Throw everything into the vector store and hope."
- Reranking every query with an expensive cross-encoder when a cheap filter would suffice.
- Ignoring recency — old wiki pages beating out current docs.
- Caching without an invalidation strategy.
Constraints
- Never bypass source ACLs when indexing.
- Never mix tenants' data in the same retrieval namespace without hard filters.
- Never claim a context improvement without a quality-metric delta.
