Session Start Protocol
Run these steps at the beginning of every session, before any task work.
1. Load directives
Call mcp__hindsight__recall with:
- bank_id:
adam-global - query:
durable directives preferences workflow defaults coding standards model strategy
Then call mcp__hindsight__recall again with:
- bank_id: the current derived
project-*orworkspace-*bank - query:
repository conventions architecture commands failures patterns
Read all results. The global bank carries durable user/workflow rules; the project/workspace bank carries repository-specific context.
2. Load mental models
Call mcp__hindsight__recall with:
- bank_id: the current derived
project-*orworkspace-*bank - query:
mental model primitives agents workflows integrations architecture
Read all results. This is your architectural map.
3. Recall task-specific context (if task is known)
Call mcp__hindsight__recall with:
- bank_id: the current derived
project-*orworkspace-*bank - query:
<the specific module, feature, or component you are about to work on>
4. Orient with CGC
Call mcp__codegraphcontext__get_repository_stats to confirm the graph is current.
If the task involves specific files/functions, also call mcp__codegraphcontext__find_code
for the target to understand its position in the graph.
5. Acknowledge
State briefly what you loaded — directives active, mental models loaded, current graph stats — so the user knows the session started warm.
Session End: Retain
Before ending the session or after completing a significant task, call
mcp__hindsight__retain with the appropriate bank and content covering:
adam-globalfor durable cross-project preferences or reusable workflow patterns- the current derived
project-*orworkspace-*bank for repository-specific decisions, failures, commands, and architectural insights
Use this format for the content:
[type: decision|pattern|failure|insight] <what happened>
Rationale/why: <the reason>
Context: <module or area affected>