second-brain

second-brain

The loop-system memory discipline — layered retrieve → work → distill — decoupled from loop machinery, for one-shot tasks. The graph (mindgap) is the hub; every substantive task starts by pulling what's already known and ends by paying its learnings back. The returns compound. All graph writes obey the ingest protocol in AGENTS.md — read it before writing.

Dispatch — do this first

• Inside a loop-system or meta-trainer session → stop. Their RESUME/SESSION END already run this discipline; doubling it corrupts the loop's memory bookkeeping.
• Task needs multi-session persistent state or a measurable done-gate → hand off to loop-system. Never imitate loop machinery. No per-task state files, ever — a task that seems to need one is a loop wearing a disguise.
• Mining asked for directly ("find connections", "what next to learn", "enrich X from my notes") → references/mining.md, skip the cycle.
• Otherwise run the cycle: RETRIEVE → WORK → UPDATE.

Stores — ~/second_brain/ (visible) + the graph (hidden)

~/second_brain/
  inbox/   drop zone — inbox-to-mindmap's default target; swept at UPDATE
  raw/     immutable archive — ingest MOVES files here (YYYY/topic/);
           source nodes carry url → the archived path
  ctx/     reusable prompts/templates/rules, each indexed by a graph node

Scaffold missing dirs on first use (mkdir -p). The DB stays in ~/.mindgap/ and is never touched by hand. Deliverables do NOT get an output/ dir — they stay in their repos; the graph indexes them (see UPDATE).

RETRIEVE (task start) — protocol: references/retrieval.md

Five layers, top-down: global learnings → project ledger → task-adjacent loop STATEs → graph (mine enrich) → ctx/. Context diet throughout: inline what is relevant, never everything. Bump used:/last: on any global row that informs a decision.