session-start

Session Start - Context Loading

Load relevant memories to maintain continuity across sessions.

When to Invoke

AUTOMATICALLY at the beginning of every session:

• When user sends their first message
• Before responding to any request
• Silently in the background

Execution Steps

1. Identify context signals:
2. Current working directory (from system context)
3. Project name (from path or CLAUDE.md)
4. User's first message keywords
5. Git branch if available

1. Search for relevant memories:

`` mcp__memory-service__memory_search( query="<project name> OR <key topic>", limit=5, time_expr="last 7 days", quality_boost=0.3 ) ``

1. Also search for:
2. Recent session summaries: query="session-summary", limit=3
3. Client info if mentioned: query="<client name>", tags=["client"]
4. Active decisions: query="decision", time_expr="last 14 days"

1. Integrate context:
2. Use loaded memories to inform responses
3. Reference relevant past decisions
4. Continue where previous sessions left off

1. Silent operation:
2. Do NOT announce "I loaded X memories"
3. Just use the context naturally
4. Only mention if user asks about previous sessions

Search Strategies

By Project

memory_search(query="botsniper trading", limit=5)
memory_search(query="foodshot ai", limit=5)

By Recency

memory_search(time_expr="last 3 days", limit=10)
memory_search(time_expr="yesterday", limit=5)

By Type

memory_search(query="decision", tags=["decision"], limit=5)
memory_search(query="client", tags=["client"], limit=5)

Combined

memory_search(
  query="<project> decisions",
  time_expr="last 7 days",
  quality_boost=0.3,
  limit=10
)

What to Look For

| Memory Type | Why It Matters | |-------------|----------------| | Session summaries | What happened last time | | Decisions | Active choices still relevant | | Open items | Unfinished work to continue | | Client info | Key details to remember | | Gotchas | Pitfalls to avoid | | Patterns | Established conventions |

Example Flow

User starts session in /Users/maskedhunter/coding/botsniper-optimizer:

1. Detect: Project is "botsniper-optimizer", domain is "trading bots"
2. Search: memory_search(query="botsniper trading bot", limit=5, time_expr="last 7 days")
3. Load: Recent decisions about Billy V4, Blood V5 parameters
4. Search: memory_search(query="session-summary", limit=2)
5. Load: Yesterday's session about optimizing stop losses
6. Respond: Use this context to continue naturally

User says: "Let's continue working on the Hurricane bot"

1. Detect: Topic is "Hurricane bot", domain is "Kalshi weather trading"
2. Search: memory_search(query="hurricane kalshi weather", limit=5)
3. Load: Recent decisions about forecast verification, position monitoring
4. Continue: Pick up where previous session left off

Cloud Recall Fallback

If local memory search returns fewer than 3 results AND cloud backup is configured (~/.claude-memory-cloud.env exists):

1. Search cloud backup:

``bash cd ~/coding/claude-memory && python3 -m cloud.cli search "<project or topic>" --limit 5 --include-deleted ``

1. If cloud has memories not found locally:
2. Include cloud results in context with [CLOUD] indicator
3. Offer to restore if user asks about missing context

1. Restore if needed:

``bash cd ~/coding/claude-memory && python3 -m cloud.cli restore --hash <hash1>,<hash2> ``

The cloud preserves everything forever - even memories that were deleted or compressed during local consolidation. This ensures total recall.

Important Notes

• Be proactive: Don't wait for user to ask for context
• Be silent: Don't announce memory loading
• Be selective: Quality over quantity - top 5-10 memories
• Be natural: Weave context into responses seamlessly
• Cloud fallback: Use cloud when local results are sparse