agentdb-memory

L1 Improvement

• Rewrote the memory guidance into Skill Forge required sections with explicit contracts and validation steps.
• Added prompt-architect constraint capture, confidence ceilings, and safety controls for data retention.

STANDARD OPERATING PROCEDURE

Purpose

Design persistent memory strategies with AgentDB, covering namespaces, retention policies, retrieval contracts, and safety/privacy requirements.

Trigger Conditions

• Positive: implementing long-term memory, audit trails, or context recall for agents/workflows using AgentDB.
• Negative/reroute: ephemeral caches or non-AgentDB storage solutions; vector search tuning (agentdb-vector-search).

Guardrails

• Define retention, encryption, and access controls before enabling writes.
• Separate memory namespaces per project/session; tag writes for traceability.
• Include data minimization and deletion workflows; avoid storing secrets in plain text.
• Maintain English outputs with explicit confidence ceilings.

Execution Phases

1. Planning: Capture use case, sensitivity, retention needs, and constraints; classify HARD/SOFT/INFERRED.
2. Schema & Namespaces: Define record schema, namespace patterns, tags (WHO/WHY/WHEN/PROJECT), and indexing needs.
3. Write/Read Paths: Specify APIs for writes, retrieval, and pruning; include rate/size limits and error handling.
4. Validation: Test CRUD operations, access controls, and retention enforcement; log results with ceilings.
5. Operations: Document monitoring, backup/restore, and incident response for memory misuse.

Pattern Recognition

• Conversation memory → chunk + summarize with time decay.
• Audit/history → append-only with strong access control and integrity checks.
• Feature store → schema evolution and validation pipelines.

Advanced Techniques

• Use tiered storage (hot/warm/cold) with TTL policies.
• Apply summarization to reduce footprint while preserving evidence.
• Add anomaly detection on access patterns for security.

Common Anti-Patterns

• Storing sensitive data without encryption or retention limits.
• Mixing unrelated contexts in one namespace causing leakage.
• No deletion/rotation plan.

Practical Guidelines

• Standardize tags: WHO=agentdb-memory-{session}, WHY=skill-execution, WHEN=timestamp, PROJECT=name.
• Document maximum record sizes and throttling behavior.
• Provide fallback behavior when reads miss (e.g., regenerate or request input).

Cross-Skill Coordination

• Upstream: prompt-architect for clarity on memory scope; skill-builder for scaffolding.
• Parallel: agentdb-vector-search for retrieval, agentdb-optimization for performance.
• Downstream: agent-creator/agent-selector using memory configs; recursive-improvement to refine retention.

MCP Requirements

• Requires AgentDB memory MCP with proper credentials and permissions; tag writes as above for auditability.

Input/Output Contracts

inputs:
  use_case: string  # required
  sensitivity: string  # required data classification
  retention: string  # required policy
  constraints: list[string]  # optional
outputs:
  memory_plan: file  # schema, namespace, and policy definitions
  validation_report: file  # access/retention tests
  runbook: summary  # monitoring, backup, and deletion steps

Recursive Improvement

• Feed incidents or retrieval misses into recursive-improvement to adjust schemas, retention, or access controls.

Examples

• Configure long-term memory for customer support agents with redaction and TTL policies.
• Set up audit-friendly memory for deployment history with access controls and backups.

Troubleshooting

• Retrieval misses → verify namespaces/tags, reindex, or adjust queries.
• Storage bloat → enable TTL, summarize, or archive cold data.
• Access issues → audit permissions and rotate credentials.

Completion Verification

• Schema, namespaces, and tagging defined.
• Retention, encryption, and deletion policies documented and tested.
• Confidence ceilings stated for reliability/safety claims.
• Runbook provided for monitoring and incidents.

Confidence: 0.70 (ceiling: inference 0.70) - AgentDB memory SOP rewritten with Skill Forge cadence and prompt-architect ceilings.