Agent Memory Conflict & Staleness Resolver

Agent memory logs often accumulate conflicting instructions and stale facts that degrade performance. This skill audits supplied memory records for contradictions, superseded data, and weak provenance to ensure the context remains accurate.

What it does

• Conflict detection identifies direct contradictions, scope overlaps, and temporal updates between memory records.
• Staleness assessment evaluates record freshness based on timestamps, volatility, and task sensitivity.
• Provenance evaluation weights records by authority, directness, and corroboration rather than just recency.
• Resolution mapping recommends specific actions including merge, qualify, quarantine, or archive for every record.
• Context packaging generates a clean, resolved context pack while flagging unresolved questions for human review.

How it works

1. Inventory all supplied memory records while preserving original source text and metadata.
2. Extract core claims, entities, dates, and conditional scopes from the inventory.
3. Analyze for duplicates and contradictions using a multi-factor freshness and authority model.
4. Report findings using the standard schema, providing evidence-based recommendations for each conflict.

Frameworks & tools

Works with any agentic framework that exports memory as text, JSON, or markdown logs. Compatible with LangChain, AutoGPT, and custom RAG implementations.

Why this beats prompting it yourself

Generic prompts often default to "most recent wins," which deletes authoritative historical data. This skill uses a nuanced logic that respects scope differences and provenance, preventing the silent loss of critical constraints.

Use cases

• Auditing long-running agent logs to remove contradictory user preferences.
• Synchronizing memory between two agents that have developed diverging facts about a project.
• Cleaning up RAG context windows by archiving superseded technical specifications.

Known limitations

Evaluates supplied records only. Cannot independently verify external truth without provided authoritative evidence. Missing timestamps result in unknown freshness rather than assumed invalidity.