LocalNest MCP Expert

Master the art of local AI engineering with LocalNest. This skill enables agents to maintain persistent intelligence across sessions, navigate vast codebases with semantic precision, and model complex relationships through a temporal knowledge graph.

📋 Mandatory SOP for AI Agents

To utilize LocalNest effectively, follow these procedural rules:

1. [COLD START] Task Initialization:

   • Rule: ALWAYS call localnest_agent_prime as your very first action for any new task.
   • Reason: This rehydrates your project context and surfaces previous decisions/standardized rules via teach.

2. [REAL-TIME] Decision Capture:

   • Rule: When you find a bug root cause or choose an architectural pattern, call localnest_capture_outcome immediately.
   • Reason: Prevents loss of context if the session resets or another agent takes over.

3. [LEARNING] Corrective Feedback:

   • Rule: If the user corrects your code style or logic pattern, use localnest_teach to store that preference.
   • Reason: Ensures you don't repeat the mistake in future tasks.

4. [COMPLETION] Task Wrap-up:

   • Rule: Call localnest_capture_outcome with status: "completed" before ending the task.

---

Core Concepts

1. Context Rehydration (Agent Prime)

Instead of repetitive searching, use Agent Prime. It performs a fused retrieval of relevant memories, knowledge graph entities, and file changes in a single call. This "rehydrates" your mental model of the project instantly.

2. Temporal Knowledge Graph (KG)

Facts aren't just strings; they are subject-predicate-object triples with a time dimension. LocalNest tracks when facts become valid and when they are superseded, allowing you to query the state of the project "as of" a specific date.

3. Proactive Memory Hints

LocalNest integrates with your file reads. When you read_file, it automatically checks for linked memories with high importance and surfaces them as hints. This prevents you from repeating past mistakes or missing documented architectural decisions.

Code Examples

Example 1: Full Task Initialization

The "Golden Path" for starting any non-trivial task.

// Rehydrate context for the current task
const context = await localnest_agent_prime({
  task: "Implement OAuth2 flow for the mobile app",
  project_path: "/abs/path/to/project"
});

// context returns:
// - relevant_memories: past decisions about auth
// - kg_entities: AuthService, TokenStore, etc.
// - suggested_actions: "Update memory id: 452 (Auth architecture)"

Example 2: Batching Knowledge Extraction

Extracting multiple facts from a code review session efficiently.

await localnest_kg_add_triples_batch({
  triples: [
    { subject_name: "AuthService", predicate: "uses", object_name: "JWT" },
    { subject_name: "JWT", predicate: "expires_in", object_name: "1hour" },
    { subject_name: "TokenStore", predicate: "depends_on", object_name: "Redis" }
  ],
  response_format: "minimal"
});

Example 3: Temporal Querying

Investigating a regression by checking the state of a component last week.

const pastState = await localnest_kg_as_of({
  entity_id: "api_config",
  as_of_date: "2026-04-01T12:00:00Z"
});

Best Practices

1. Prefer agent_prime over search_hybrid: It is significantly more token-efficient for re-establishing context.
2. Use Batch Tools for 3+ Items: Tools like memory_store_batch use a single database transaction, ensuring consistency and massive speed gains.
3. Capture Outcomes: Always call capture_outcome or memory_store after a major decision. Memory is the "learned behavior" of your agent.
4. Minimal Payloads: Pass terse: "minimal" when you don't need to read back what you just wrote.

Advanced Patterns

The "Memory-First" Workflow

1. Recall: Check for existing patterns using find.
2. Execute: Build the feature.
3. Reflect: Capture the new knowledge using capture_outcome.
4. Link: Connect the file to the memory using file_changed.

Troubleshooting

Issue: Low Search Relevance

Solution: Broaden your query, disable filters, or try find({ sources: ["memory", "triple"] }) to check if the knowledge exists in the graph but not in the code.

Issue: Knowledge Contradictions

Solution: LocalNest warns but doesn't block. Use kg_timeline to see how the fact evolved and kg_invalidate to mark the old fact as stale.

References

• LocalNest Documentation
• PCL Skill Specification