| name | session-start |
| description | Load relevant memories at session start for context continuity.
IMPORTANT: Use this AUTOMATICALLY at the START of every session.
Search for memories related to:
- Current working directory/project
- Recent activity (last 7 days)
- Topics the user mentions in their first message
Execute silently - do not announce what you're loading.
|
| user-invocable | true |
| disable-model-invocation | false |
| argument-hint | [optional topic to focus on] |
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
-
Identify context signals:
- Current working directory (from system context)
- Project name (from path or CLAUDE.md)
- User's first message keywords
- Git branch if available
-
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
)
-
Also search for:
- Recent session summaries:
query="session-summary", limit=3
- Client info if mentioned:
query="<client name>", tags=["client"]
- Active decisions:
query="decision", time_expr="last 14 days"
-
Integrate context:
- Use loaded memories to inform responses
- Reference relevant past decisions
- Continue where previous sessions left off
-
Silent operation:
- Do NOT announce "I loaded X memories"
- Just use the context naturally
- 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:
- Detect: Project is "botsniper-optimizer", domain is "trading bots"
- Search:
memory_search(query="botsniper trading bot", limit=5, time_expr="last 7 days")
- Load: Recent decisions about Billy V4, Blood V5 parameters
- Search:
memory_search(query="session-summary", limit=2)
- Load: Yesterday's session about optimizing stop losses
- Respond: Use this context to continue naturally
User says: "Let's continue working on the Hurricane bot"
- Detect: Topic is "Hurricane bot", domain is "Kalshi weather trading"
- Search:
memory_search(query="hurricane kalshi weather", limit=5)
- Load: Recent decisions about forecast verification, position monitoring
- 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):
-
Search cloud backup:
cd ~/coding/claude-memory && python3 -m cloud.cli search "<project or topic>" --limit 5 --include-deleted
-
If cloud has memories not found locally:
- Include cloud results in context with [CLOUD] indicator
- Offer to restore if user asks about missing context
-
Restore if needed:
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