xvision-memory-ops

name	xvision-memory-ops
description	Operate xvision's persistent memory layer: inspect Observations and Patterns, preview recall safely, manage Pattern lifecycle, run leakage probes, and avoid F+L+T violations.

xvision memory ops

Use this skill when operating the xvision-memory substrate from CLI, dashboard, or MCP. Memory is trading-critical state: treat every command as an audited operator action.

When to use

Run this skill when operating the memory substrate: listing or inspecting Observations and Patterns, previewing recall safely, managing Pattern lifecycle (activate / retire / undo-forget), or running leakage probes.

When NOT to use

• Distilling new Patterns or running the autoresearch gate → use xvision-autoresearch-ops
• Reading flywheel velocity, lineage, or training-run example pools → use xvision-flywheel-ops
• MCP recall used as a live write path (not preview) → disallowed; stop

Trigger examples

• List all active Patterns for agent abc123
• Show me the Observations for agent abc123 from last week
• Activate Pattern p_005 for agent abc123
• Retire Pattern p_003 — check lineage first
• Run the leakage regression script
• Preview recall for this embedding without writing to the prompt
• Undo the last forget on agent abc123
• Show me the memory-ops safety contract

Safety Contract

• Observations are concrete per-cycle records. They are never recalled into live prompts.
• Patterns are distilled or operator-attested semantic memories. Only active, non-forgotten Patterns can be recalled.
• Backtests must pass a scenario start. Recall must exclude Patterns whose training_window_end overlaps that scenario.
• Case-law framing is required whenever Patterns enter a prompt.

CLI Checks

xvn memory ls --agent <agent_id> --tier pattern --json
xvn memory ls --agent <agent_id> --tier observation --json
xvn flywheel status --agent <agent_id> --json
bash scripts/leakage-regression.sh

For lifecycle work:

xvn memory activate <pattern_id> --json
xvn memory retire <pattern_id> --json
xvn memory undo-forget --agent <agent_id> --json

MCP Read Tools

• xvn_memory_list
• xvn_memory_get
• xvn_memory_recall
• xvn_flywheel_status
• xvn_flywheel_velocity

MCP recall requires a caller-supplied embedding. It must be treated as preview or agent-assist context, not as a write path.

Evidence To Capture

• CLI transcript for list/recall/lifecycle operations.
• Leakage-regression output after changes touching recall, query, prompt rendering, Pattern lifecycle, or provenance.
• Before/after JSON when deleting, retiring, activating, or restoring rows.

Gotchas

Recalling Observations into live prompts: Observations are never recalled into live prompts — only Patterns can be recalled. Doing otherwise violates the safety contract.

Missing scenario start in backtests: Recall must exclude Patterns whose training_window_end overlaps the scenario start. Failing to pass the scenario date to the recall command leaks future data.

MCP recall as write path: MCP recall is preview or agent-assist context only. Using its output as a direct prompt write path bypasses the case-law framing requirement.

Activating without case-law framing: Patterns entering a prompt require case-law framing. Activation alone does not satisfy this — the recall layer must apply the framing at query time.

Skipping leakage regression: After any change to recall, query, prompt rendering, Pattern lifecycle, or provenance, the leakage-regression script must run. Omitting it leaves F+L+T violations undetected.

Hard delete without evidence capture: Before deleting or retiring any row, capture the before/after JSON. The evidence section is not optional for lifecycle operations.

Owner

memory-ops track (team/contracts/memory-ops.md)