2nd Brain

Preamble

2nd Brain asks when memory earns authority. It stores lessons, decisions, source-backed facts, and chat residue only so they can be challenged before they steer the next plan, spec, prompt, skill, gate, or evaluator.

---

Memory Starts As Authority

A second brain sounds like a place to remember more. The danger is that remembered context starts giving orders.

An old note can look wise because it survived. A generated summary can look settled because it is fluent. A retrieved decision can keep shaping work after the evidence that made it true has expired. That is the pressure behind this project. Memory has to decay, compete, cite itself, and earn authority again each time it tries to guide future work.

The Wiki That Wants To Compound

Karpathy’s LLM Wiki gives the first shape: read sources, compile knowledge into a persistent wiki, and file useful answers back into the system. The promise is compounding. Work done once can make the next answer sharper.

Scale turns the promise into a maintenance problem. More pages means more stale claims, contradictions, orphaned sources, and smooth summaries that future agents may treat as fact. A wiki can become a hallucination with backlinks. What carries over is the loop: source, synthesis, retrieval, correction, writeback.

A Brain With Gates

Underneath the metaphor, this is a hybrid retrieval system: graph edges, vector search, and lexical search across SQLite and Obsidian, with embeddings generated through Modal and Ollama where semantic recall is useful.

The vector index is a candidate finder. Authority comes from source role, provenance, freshness, contradiction handling, review status, and whether the retrieved lesson improves a future surface.

The spine keeps source families separate. Business doctrine, project state, personal memory, therapist material, logbook material, morning pages, chat transcripts, PRDs, specs, and generated drafts enter with different authority. Raw material enters as artifact. Extraction creates candidates. Review decides what survives.

The Loop That Learns

Spark Researcher adds the recursive question: how does the brain get better at getting better?

The useful pattern is small and strict: run a command, read a fixed metric, write the ledger, save the lesson, suggest the next bounded experiment, and prepare reviewable changes. 2nd Brain adapts that loop to agent work.

agent work
  -> retrieve bounded context
  -> produce output
  -> run gate or score
  -> record failure or improvement
  -> extract lesson
  -> review/apply
  -> patch prompt, skill, gate, evaluator, or template
  -> next agent starts with a better floor

The system may propose improvements. It may prepare patches. It may show which lesson created which change. Human review controls rule changes future agents must obey.

Chats Become Evidence

Some of the most useful memory lives in conversations: corrections, preferences, rejected approaches, repeated failures, decisions, and the moment a piece of work finally lands.

Claude Memory Compiler points at the capture pattern: session logs, pre-compact saves, extracted decisions, lessons, gotchas, and structured knowledge. 2nd Brain makes the boundary stricter. A chat transcript is raw evidence. A memory extracted from it is a candidate. A durable rule needs review. This is how judgment gets written back into the system without turning every frustrated comment into doctrine.

The Cost Of Remembering

PAI and OB1 point toward a broader personal AI layer: skills, hooks, agents, memory, identity, one database, one gateway, many clients. The temptation is to make the brain bigger until it feels complete.

The hard work is smaller and less glamorous: delete, demote, expire, separate, score, cite, and refuse. I decide which lesson deserves to survive. The machine makes that decision cheaper to inspect and harder to fake.

When Proof Changes Shape

Evidence Garden is the visual consequence of the same argument. Links, recent activity, and convincing summaries count only when the evidence trail gives them weight. A claim should thicken, dim, branch, or crack because provenance, confidence, freshness, and contradiction changed. Beauty is suspect until the evidence trail explains the shape.

Proof Of Landing

This project has launched locally through the Memory runtime. Governed query and supervised compile paths already run from bounded packs with gates.

Landing is harder. It means accepted doctrine deltas repeatedly change future prompts, skills, gates, evaluators, PRD/spec templates, and plans. It means smaller context packs, lower token cost, fewer broad reads, fewer repeated mistakes, and less rework across real projects.

The public garden can show the premise, architecture, metrics, learnings, anti-patterns, and selected snippets. Proprietary playbooks, private decisions, raw source material, and unreviewed doctrine stay local.