Architecture

Knowledge base MCP tools

Agents reach the knowledge base through the Model Context Protocol — 13 tools across four jobs: priming context, searching, learning from failure, and remembering across sessions.

Context and search

Tool	What it does
`kb_context`	The first call of any task: decomposes the request into optimized queries, retrieves across all tiers in parallel, returns a synthesized briefing (~1,500 tokens instead of 6,000+)
`kb_search`	Narrow follow-up queries on a specific topic after initial context is loaded
`kb_status`	Health and index state of the knowledge base

Ingestion and outputs

Tool	What it does
`kb_ingest`	Adds new documents to the knowledge base
`kb_submit_output`	Records project output data (Tier 3) for later evaluation
`kb_submit_outcome`	Records how an output performed, closing the optimization loop

Error memory and learnings — the self-annealing surface

Tool	What it does
`kb_check_errors`	Called BEFORE risky operations (deploys, migrations, bulk sends): returns known errors matching the planned action, with root causes and prevention rules
`kb_log_error`	Called AFTER resolving any non-trivial failure: severity (S1 prod-down / S2 blocked / S3 friction) plus exactly one resolution — prevention-rule, check-added, or accepted-risk
`kb_update_learning`	Appends durable discoveries to the relevant directive’s Learnings section

Session memory

Tool	What it does
`memory_save`	Persists a discovery for future sessions
`memory_search`	Recalls relevant past work
`memory_timeline`	Reviews session history chronologically

Vault writes

Tool	What it does
`vault_write`	Files notes into the bot-writable vault areas: errors, learnings, sessions, reviews, and inbox

The two-lane rule

Writes follow two lanes. Lane A (automatic): errors, learnings, sessions, and reviews go straight to the vault through the tools above. Lane B (reviewed): changes to directives and SOPs — the normative documents — go through a pull request in the owning repository, reviewed by a human. Agents record what happened freely; they don’t rewrite the rules without review.

Next steps

Knowledge base overview — the three-tier system these tools serve
The DOE framework — where agents sit in the architecture