A working personal methodology for grounded retrieval, observable iteration, context recovery, and human-approved output.
My Second Brain combines a local knowledge vault with a hill-climbing loop runner. It retrieves relevant decisions and source material, produces a candidate, evaluates it across explicit dimensions, compares the result with the previous version, restores anything important that was lost, and keeps only changes that move the work forward.
A useful Second Brain needs two connected systems: a memory path that retrieves grounded context and an optimization path that can improve output without silently dropping intent.
Files remain human-readable and organized by domain. Each item carries enough metadata to trace its origin, recency, project, and intended use.
vault/{notes, decisions, projects, clients}The first four areas reflect the repository structure demonstrated on the live implementation. The run-history folder is the loop-engineering extension that makes improvement observable and recoverable.
Reusable commands, agent configuration, hooks, guardrails, evaluation rules, and output contracts.
commands/proposal.mdcommands/brief.mdhooks/agents.jsonevaluation-rubric.jsonThis fixed replay reflects the working methodology described in my loop essay. It demonstrates how changes and regressions are recorded. The rubric scores belong to this single documented run; they are not general model benchmarks.
Tightened the final sections and restored missing nuance.
Accuracy, flow, and completeness.
Marginal gain indicated convergence.
A candidate is rejected if it drops a required fact, source boundary, audience need, or explicit instruction—even when its total score rises.
Clarity cannot compensate for factual inconsistency. Each critical dimension has its own threshold and regression check.
Every candidate, diff, scorecard, and decision remains available so the system can restore the last known good state.
The run stops when improvement becomes marginal, a quality target is met, or the allowed time, token, or cost budget is exhausted.
Retrieved context keeps its file reference. Unsupported claims are flagged for review rather than converted into confident prose.
The loop can recommend a candidate. A person remains accountable for sensitive decisions, publication, and external action.
Retrieve prior theses and decisions, improve the argument over multiple iterations, and keep a visible record of what changed.
Connect material across sources, preserve disagreement, flag evidence gaps, and separate cited fact from interpretation.
Improve clarity, structure, tone, and claim discipline while protecting the original positioning and constraints.
Turn notes into decision-ready briefs while retaining risks, assumptions, alternatives, and accountable next actions.
Recover earlier trade-offs and design rationale before evaluating a new option or writing an ADR.
Synthesize recent logs into status updates while preserving continuity across weeks and projects.
Choose a strategy note, research synthesis, proposal, brief, or weekly update with a clear definition of good.
Place human-readable material in the vault and distinguish source documents from generated artifacts.
Chunk and embed locally while preserving file path, timestamps, project metadata, and stable source identifiers.
Define quality dimensions, must-keep constraints, failure conditions, and the stopping budget before generation begins.
Store the prompt, retrieved context references, output, scorecard, diff, trade-offs, and keep-or-retry decision.
Treat the selected result as a reviewed artifact—not an autonomous truth or an unbounded action.
These sources validate the underlying patterns: iterative refinement, reflective memory, evaluator–optimizer loops, long-term memory, checkpointing, local retrieval, and metric-driven optimization. They do not independently validate the specific scores in the public-safe run replay.
Iterative feedback and refinement can improve outputs without additional model training.
Agents can retain verbal reflections in episodic memory and use them to improve subsequent trials.
Documents the evaluator–optimizer workflow: one model call generates while another evaluates and feeds back in a loop.
Explains multi-dimensional evaluation, transcript inspection, and the need to catch regressions before deployment.
Documents long-term memory, semantic search, checkpoints, replay, human review, and recovery.
Formalizes metric-driven optimization of prompts and LM programs, including coordinate-ascent and reflective optimizers.
Provides a persistent client that stores vector data on disk for local development and testing.
Documents local embeddings for semantic search, retrieval, and RAG pipelines.
This is an evolving personal methodology for building better knowledge work with local context, explicit evaluation, recoverable versions, and human judgment.
For conversations about local-first memory, retrieval quality, evaluation loops, and human-controlled AI workflows.