Skip to main content
Most knowledge systems are write-once: you dump content in, you query it out, it degrades silently as it grows. Brain is different — it runs a closed loop that continuously measures, cleans, links, and tunes itself. The entry point is a single command: brain auto-kb.

​
The ACE pattern

Brain implements the Autonomous Context Engineering pattern adapted from recent agent research: 
Generate ──▶ Reflect ──▶ Curate ──▶ Evolve ──▶ (loop)
PhaseWhat happensCommands involved
GenerateSearch with confidence scoring, authority weights, link boostbrain search, brain analyze
ReflectMark results helpful / harmful, update source reputation, track search quality per domainbrain reflect
CurateDomain routing, semantic dedup, health scoring, auto-link generationbrain dedup, brain link, brain health
EvolveParameter sweeps, strategy segment tuning, graph-structure analysisbrain sweep, brain strategy, brain graph
Each phase writes back to the brain vault as markdown under _reflections/, _decisions/, and _domains/ — so everything the optimizer learns is auditable and portable. Nothing is hidden in a database row.

​
Running it

The one-liner is the autonomous loop: 
brain auto-kb --brain my-brain --rounds 3
That runs 3 rounds of ACE: generate (search with a seeded question set) → reflect (grade the answers) → curate (dedup, relink, health-fix) → evolve (sweep retrieval params, persist the winning strategy). Every round produces a markdown audit trail in your vault.

​
What auto-kb actually does per round

1

Probes retrieval with a curated question set

Either the defaults shipped with Brain, or a list you provide via --questions questions.md.
2

Scores every answer against authority + confidence

4-dimension quality grade per question. Bad answers flagged for follow-up curation.
3

Curates the vault

  • brain dedup --threshold 0.85 — collapse near-duplicates (with audit).
  • brain link — generate tag-overlap and semantic links; boost graph connectivity.
  • brain health — lint the vault, fix obvious issues, re-embed stale chunks.
4

Sweeps retrieval parameters

Grid search over RRF k, link-boost cap, source-type weights; picks the setting that wins against the question set.
5

Persists decisions

The winning strategy is saved to _decisions/ and applied via brain strategy set. Rolled back automatically if the next round regresses.

​
The individual commands

If you want to drive the loop manually instead of via auto-kb, each phase has its own command: 
brain reflect --brain my-brain
# Flags results helpful/harmful, updates source reputation

​
Why this matters

Most “agent memory” solutions are vector stores with a nice wrapper. They grow; they don’t improve. After six months you have 40,000 chunks and no sense of which ones are actually useful. Brain’s closed loop gives you:
  • A quality floor. brain score returns a grade. If it drops, you know before your agent starts giving bad answers.
  • Self-pruning. brain dedup catches the copy-paste-from-Slack patterns that bloat vector stores.
  • Link uplift. brain link uplifts search scores for docs that relate to confirmed-good answers — so the good stuff stays findable as the vault grows.
  • Tuning without guessing. brain sweep replaces “is k=60 the right RRF constant?” with an empirical answer against your actual content and questions.

​
Audit trail

Everything auto-kb does lands in the vault: 
my-brain/
├── _decisions/
│   └── 2026-04-17-strategy-segment-code-queries.md    ← winning sweep
├── _reflections/
│   └── 2026-04-17-run-42.md                           ← question set + grades
├── _domains/
│   └── code.md                                         ← routing config
└── logs/
    └── auto-kb-2026-04-17.md                          ← run summary
You can open any of these in Obsidian, review what changed, roll back by reverting the markdown file, and commit the next run to git. The optimizer is not a black box — it writes its reasoning down.

​
What’s next

Knowledge graph operations

The graph ops that make the closed loop possible — link, dedup, health, score.

Configuration

Deployment modes, env vars, and the brain.json strategy block.