context-loader

Context Loader

Load relevant memories from Pensyve at the start of a session to provide cross-session continuity. Respects the context_loading configuration setting.

Instructions

Note on invocation: The SessionStart hook has already performed a scoped recall and (if applicable) injected a continuity primer. Invoking this skill adds deeper on-demand context — not a replacement for the hook's work. Expect to see both.

When this skill is invoked (typically at session start), follow these steps:

Step 0: Determine Loading Mode

Check the mode argument if provided. Otherwise, respect the context_loading setting from the plugin configuration (pensyve-plugin.local.md):

• off: Do nothing. Inform the user: "Context loading is disabled. Use /recall to search memories on demand."
• summary: Load a concise overview (10-15 lines max).
• full: Load comprehensive context with relevance scores and details.

If no mode is specified and no configuration is found, default to summary.

Step 1: Query Memories

Run the following pensyve_recall queries to gather session context:

1. Recent decisions: pensyve_recall with query "decided" (limit: 5)
2. Known issues: pensyve_recall with query "issue OR bug OR error OR problem" (limit: 5)
3. Workflow patterns: pensyve_recall with query "workflow OR pattern OR process" (limit: 5)
4. Recent activity: pensyve_recall with query "*" (limit: 10) -- broad query to capture recent memories by recency

Deduplicate results across queries (same memory ID should appear only once).

Continuity-aware primer

If the SessionStart hook detected a continuation (check plugin session state for recent_context — this is set by the hook when shared-entity score ≥0.7 against episodic recall results; it is a plugin-layer concept, not a structured server-side episode link), structure the primer around continuity rather than fresh recall:

Continuing: <entity-set> — prior work context

Last lessons:

• [recent observation 1]
• [recent observation 2]

Open questions carried forward:

• [if any observations had open-question provenance]

Note: Continuity is best-effort — it is inferred from shared-entity overlap in recent episodic recall, not from a persisted episode-to-episode link. The MCP server has no episode-listing API today. When in doubt, the hook falls back to a fresh-session primer.

If no continuation was detected, present a standard recall-based primer.

Step 2: Present Context

Summary Mode (10-15 lines max)

Present a concise briefing with the most important items:

Session Context (from Pensyve memory)

Recent Decisions:

• auth-service: Using RS256 for JWT signing to support key rotation
• api-design: POST endpoints return 201 with created resource

Known Issues:

• database: Migration script requires Python 3.11+

Active Patterns:

• testing: Integration tests need tmpdir cleanup after filesystem operations

Use /recall <query> to search for specific memories.

Rules for summary mode:

• Maximum 15 lines of content (excluding the header)
• Show at most 3 items per category
• Omit categories with no results
• Do not show scores, IDs, or timestamps
• Prioritize higher-confidence and more recent memories

Full Mode (comprehensive)

Present a detailed briefing with scores and metadata:

Session Context (from Pensyve memory)

Recent Decisions (3 found):

Entity	Decision	Confidence	When
auth-service	Using RS256 for JWT signing	0.9	2026-03-15
api-design	POST endpoints return 201	0.9	2026-03-14
database	SQLite for MVP, migrate to Postgres later	0.9	2026-03-12

Known Issues (2 found):

Entity	Issue	Confidence	Score
database	Migration requires Python 3.11+	0.8	0.91
cache	Invalidation race condition on concurrent writes	0.8	0.85

Workflow Patterns (1 found):

Entity	Pattern	Confidence	Score
testing	Integration tests need tmpdir cleanup	0.7	0.78

Recent Activity (5 unique memories, most recent first):

Type	Entity	Summary	Score
semantic	auth-service	RS256 JWT signing	0.92
episodic	database	Debugged migration failure	0.88
procedural	deploy	Run tests before deploy	0.75

Total memories loaded: 11 | Use /recall <query> for targeted search | /inspect <entity> for entity details

Rules for full mode:

• Show all results with relevance scores
• Include confidence values and timestamps where available
• Group by category with counts
• Show memory types in the recent activity section
• Include a footer with total count and navigation suggestions

Constraints

• Never read or write .claude/ memory files. All memory operations go through the Pensyve MCP tools exclusively.
• Do not fabricate memories. Only display what the MCP tools return.
• In summary mode, strictly respect the 15-line content limit. Truncate rather than exceed it.
• Do not take any action based on the loaded context -- this skill is informational only.
• If no memories are found at all, say: "No memories found in the current namespace. This appears to be a fresh session. Use /remember to start building context."
• The context loading should be fast and non-blocking. Do not run expensive operations.

Error Handling

• If some pensyve_recall queries fail but others succeed, present the successful results and note the failures briefly.
• If all queries fail, report the error and suggest checking the MCP server connection.
• If the MCP server is not connected, inform the user: "Pensyve MCP server is not connected. Context loading skipped. Verify your MCP server configuration."