memory-review

name	memory-review
description	Memory hygiene audit — checks for stale, contradictory, low-confidence, and consolidation-candidate memories, then offers cleanup with user confirmation

Memory Review

Audit Pensyve memory health and identify issues that need attention: stale memories, contradictions, low-confidence entries, and consolidation candidates.

Instructions

When this skill is invoked, follow these steps:

Step 1: Gather Memory Inventory

If an entity is specified, call pensyve_inspect with that entity (limit: 50). If no entity is specified, run a broad pensyve_recall query with "*" (limit: 50) to discover entities, then inspect the top entities found.

Collect all memories for analysis.

Step 2: Run Hygiene Checks

Analyze the collected memories for four categories of issues:

Check 1: Staleness (> 30 days unaccessed)

Identify memories where last_accessed is more than 30 days ago or where retrievability has dropped below 0.3 (indicating FSRS decay). These memories are at risk of becoming irrelevant.

Flag criteria:

• last_accessed older than 30 days from today
• retrievability below 0.3
• access_count of 0 (never retrieved since creation)

Check 2: Contradictions (conflicting predicates)

Identify semantic memories for the same entity that have conflicting predicates or objects. Look for:

• Same subject + predicate but different objects (e.g., "auth-service uses HS256" vs "auth-service uses RS256")
• Temporal conflicts where an older fact contradicts a newer one but has no invalid_at set
• Opposing outcomes in procedural memories for the same trigger/action

Check 3: Low Confidence (< 0.5)

Identify memories with confidence below 0.5. These may have been stored speculatively or may reflect uncertain information.

Flag criteria:

• confidence below 0.5
• Procedural memories with reliability below 0.5 and trial_count above 3

Check 4: Consolidation Candidates

Identify memories that could benefit from consolidation:

• Multiple episodic memories about the same topic that could be promoted to a semantic memory
• Semantic memories that reinforce each other and could be merged
• Procedural memories with high trial counts and stable reliability that are proven patterns

Step 3: Present Audit Report

Present the findings in a structured report:

Memory Hygiene Report

Audited: 42 memories across 5 entities

Stale Memories (4 found)

#	Entity	Memory	Last Accessed	Retrievability
1	old-api	Used XML responses	45 days ago	0.15
2	config	Default port was 3000	38 days ago	0.22
3	testing	Used mocha for tests	60 days ago	0.08
4	deploy	Manual deploy to staging	33 days ago	0.28

Contradictions (1 found)

#	Entity	Memory A	Memory B	Issue
5	auth-service	"uses HS256 signing" (conf: 0.8)	"uses RS256 signing" (conf: 0.9)	Conflicting signing algorithm

Low Confidence (2 found)

#	Entity	Memory	Confidence
6	cache	"might need Redis for sessions"	0.3
7	api	"possibly rate limited at 100 req/s"	0.4

Consolidation Candidates (1 found)

#	Entity	Description	Suggestion
8	database	3 episodic memories about migration fixes	Promote to semantic: "migration script requires version check"

Summary: 4 stale, 1 contradiction, 2 low-confidence, 1 consolidation candidate

If a category has no issues, omit that section entirely. If all checks pass, report:

Memory Hygiene Report

Audited: 42 memories across 5 entities

All checks passed. Memory store is healthy.

Step 4: Offer Actions

After presenting the report, offer cleanup actions with user confirmation:

Recommended Actions:

1. Archive stale memories (#1-4): Remove from active recall. Use pensyve_forget on each.
2. Resolve contradiction (#5): Keep the RS256 memory (higher confidence, more recent). Forget the HS256 memory.
3. Review low-confidence (#6-7): Confirm or remove these uncertain memories.
4. Run consolidation (#8): Promote episodic patterns to semantic memories via pensyve_remember.

Which actions should I take? (e.g., "1,2", "all", "none")

NEVER execute actions without user confirmation. Wait for explicit approval before calling any MCP tools to modify memory.

Step 5: Execute Confirmed Actions

For confirmed actions:

• Archive/forget: Call pensyve_forget with the entity name for each confirmed deletion.
• Resolve contradiction: Call pensyve_forget for the outdated memory. If both should be kept, note the conflict and move on.
• Consolidation: Call pensyve_remember to store the promoted semantic memory, then optionally pensyve_forget to clean up the source episodic memories.

Report results after each action.

Constraints

• All memory operations go through the Pensyve MCP tools exclusively. Do not read or write to the local filesystem for memory purposes.
• NEVER delete or modify memories without explicit user confirmation. This is a hard requirement.
• Do not fabricate findings. Only report issues based on actual data from MCP tools.
• The staleness threshold is 30 days. Do not change this without user request.
• The low-confidence threshold is 0.5. Do not change this without user request.
• Entity names MUST be lowercase and hyphenated.
• When checking for contradictions, only flag clear conflicts -- do not flag complementary facts as contradictions.
• The audit should be informational first, actionable second. Present the report before offering actions.

Error Handling

• If pensyve_inspect fails for an entity, skip it and note the failure in the report.
• If pensyve_recall returns errors, report partial results and note which queries failed.
• If pensyve_forget fails during cleanup, report the error and continue with remaining actions.
• If the MCP server is not connected, inform the user and suggest checking their PENSYVE_API_KEY environment variable and MCP configuration.