.cursorrules

# MemBerry project guidance

Use the global MemBerry contract and the `$memberry` skill for non-trivial work. Repository/runtime evidence outranks recalled memory. Consolidation and full wiki publication are automatic; manual operations are verification or recovery only.

## MemBerry Memory

Project: memberry
Description: Persistent graph, retrieval, consolidation, code-intelligence, and wiki memory engine for coding and general-purpose agents.
Domain: agent memory and context infrastructure
Project Tag: project:memberry
Workload Profile: coding

Entities:
- memberry
- mcp-server
- memory-core
- consolidation-engine
- retrieval-engine
- neo4j-store
- redis-signal-queue
- wiki-compiler
- wiki-viewer
- code-intelligence
- architecture-intelligence
- graph-analytics

Canonical Entity IDs:
- memberry-project: Y-GCkJYdEeWm38j_HI1XX

Stable Tags:
- backend
- database
- infrastructure
- retrieval
- consolidation
- wiki
- coding-agents
- observability
- security

Recall Priorities:
- Active approved decisions and supersession history.
- Current architecture boundaries, interfaces, and blast radius.
- Known lifecycle failure modes and failed recovery attempts.
- Tenant/project isolation and provenance constraints.
- Readiness, publication, and indexing health relevant to the task.

Store Policy:
- Classify durable knowledge and verify every returned episode.
- Final decisions use `memory_type: decision` and `outcome: approved`.
- Pattern/convention stores are independently sourced evidence, not invented approval.
- Resolve canonical Entity IDs and reuse the stable tag vocabulary.
- Search for an existing Semantic before emitting reinforcement, correction, or contradiction.

Data Exclusions:
- Secrets, credentials, tokens, raw customer data, routine diffs, and code derivable from git.

Agent Coordination:
- Share one conversation session ID; use distinct runtime/connection `agent_id` identities for parent and child agents. Do not invent unsupported tool fields.
- Subagents load only task-relevant context and return durable findings to the parent.
- The parent stores one integrated decision or handoff instead of duplicate worker summaries.

Priors:
- The repository is a TypeScript monorepo whose runtime graph and source indexes can drift independently.
- Historical unclassified semantics may still appear through conservative wiki fallbacks.
- Live integration behavior can differ from mocked Redis and Neo4j tests.