Every new conversation starts from zero. You repeat your preferences, projects, and decisions, while useful context slowly disappears into old chats and scattered notes.

Enhanced Second Brain turns a folder of portable knowledge files into a private memory that any local AI agent can search, maintain, and keep tidy. It is provider-neutral: the same files and the same single context command work with local agents powered by OpenAI, Anthropic, Google, open models, or future models. The model can change while your context stays yours.

Send this message to any local AI agent that can access files and run commands:

Install Enhanced Second Brain from https://github.com/buffalodebile/enhanced-second-brain. Follow AGENT_INSTALL.md, use my existing notes folder if I provide one, and verify everything.

That is the human installation flow. The agent detects the machine, downloads one verified self-contained application bundle, adopts or creates the knowledge folder, builds search, connects its own persistent instructions, and runs a health check. Maintenance then happens naturally on future agent requests; no Windows task, cron job, or background service is created by default.

There is no Windows installer to click and no runtime, package manager, plugin, or database service for the human to configure. The bundle is a normal ZIP that the agent verifies and extracts; its small runtime directory is what avoids re-extracting a large single-file program on every request.

• Less repetition — previous decisions and project context can be recovered before answering.
• Faster retrieval — search opens the few useful notes instead of scanning the whole folder.
• Cleaner memory — new and edited files are indexed automatically; cold notes are archived, never deleted.
• Privacy — notes, usage signals, and search stay on the machine.
• Freedom and context sovereignty — the memory belongs to you, not to a model account. Export it as one standard ZIP, move it to another machine, open it in another editor, or connect it to another local AI.

Ask your local agent:

Export my Enhanced Second Brain as one portable bundle, verify it, and tell me where you saved it.

The resulting .zip contains the current OKF Markdown knowledge, reversibly archived notes, attachments, local settings, and usage history. Every file has a SHA-256 checksum in a small manifest. It does not contain the disposable FTS5 index, Git internals, editor state, or machine-specific agent instructions. Those are rebuilt for the destination machine.

To move it, copy that ZIP to a private drive, encrypted storage, or another computer. Then give the repository URL and the ZIP to any supported local agent and say:

Restore this Enhanced Second Brain bundle into my notes folder, reconnect automatic maintenance, and run the health check.

The agent verifies every file, rejects unsafe paths or modified content, restores into a new folder, audits OKF, rebuilds search, and regenerates local integration. Without the toolkit, the ZIP is still readable with any ordinary archive tool because the knowledge remains UTF-8 Markdown. See Context sovereignty and portable export for the exact contents, commands, migration flow, security limits, and difference between export and backup.

On September 1, 2026, we measured 60 searches across 500 OKF Markdown notes on Windows 11 with an Intel Core i7-13620H. Enhanced Second Brain found the right note every time and took about 180 milliseconds, compared with about 225 milliseconds when rereading the whole folder.

In plain English: the useful note appeared about 45 milliseconds sooner, which was about 20% less searching time in this test. More importantly, the AI was shown at most five likely notes instead of having to inspect all 500. It gets to the useful context with less irrelevant text and less work. The first setup took 3.9 seconds once; after that, only changed notes needed updating.

This does not mean every complete AI answer will be 20% faster. Writing the answer, network speed, computer speed, and the contents of the notes also take time. The result only measures how quickly the system found the right local note. See the method and exact result, then test your own notes before making a broader claim.

A same-machine scaling follow-up found 24.1% less retrieval time at 500 notes and 23.5% at 1,000. At 2,500 notes, the current safety-first freshness scan made the FTS5 path 12.9% slower than a warm full-folder scan. Enhanced Second Brain is therefore optimized for ordinary personal vaults, not claimed as the fastest possible engine at every corpus size. The exact runs and limitation are in the scaling results.

OKF does not replace Markdown. Open Knowledge Format v0.2 defines a knowledge bundle as a directory of Markdown files with YAML metadata at the top.

• Markdown is the readable file and its content: headings, paragraphs, lists, tables, and links.
• OKF is the portable contract around that content: type, title, description, sources, provenance, generation time, and lifecycle.

In short: Markdown is how the knowledge is written; OKF is how humans and machines agree on what that knowledge means. The same .md file provides both.

---
type: Decision
title: Keep search local
description: Use a disposable local index over portable knowledge files.
tags: [architecture, privacy]
sources:
  - resource: notes/architecture-review.md
generated: {by: process:enhanced-second-brain, at: 2026-09-01T00:00:00Z}
status: stable
---

# Keep search local

The knowledge files remain authoritative. The search database can always be rebuilt.

FTS5 is SQLite's full-text search engine. It breaks titles, tags, summaries, headings, and body text into searchable terms, keeps an inverted index, and ranks matches with BM25 relevance.

When a query arrives, Enhanced Second Brain checks which files are new, changed, renamed, or gone. Only that delta is reindexed; unchanged files are not reparsed. Search then returns a small ranked set for the agent to read. The index is disposable: delete it and it rebuilds from the OKF Markdown files. No GPU, embeddings, network request, or continuously running process is involved.

At the start of each request, the agent runs one local command: context. That command performs a tiny maintenance check, updates changed files in FTS5, searches the request, returns likely pages, and records which results were shown. The agent then reads only relevant pages, records those it actually uses, and distills durable new knowledge back into the right page. Raw conversation text is not blindly copied into the knowledge base.

An injected search result counts 0.25, an opened page counts 1, and a page actually used in an answer counts 2. Older activity gradually loses weight, so something that was popular years ago does not stay permanently important. Links, confidence, importance tier, freshness, frequency, and last use are combined into the page's utility score.

The first context call after 24 hours—or after 50 interactions—runs a short local reconciliation:

1. brings new or edited Markdown into the OKF profile;
2. validates required metadata and source structure;
3. synchronizes and verifies the FTS5 index;
4. recalculates page utility from real usage.

Most requests only read one small state file, so there is no heavy maintenance on every prompt. The process does not invent facts or rewrite meaning without a source and an agent judgment.

Pages become archive candidates only after at least 240 days without use, low effective usage, a minimum age, and a cold utility score. Important, verified, confidential, hub, and sufficiently linked pages are protected. On the first request after 30 days, the agent reads each candidate and its relationships before selecting what can safely move. Eligible pages go to a dated archive and can be restored; automatic deletion is never used.

Operating-system scheduling remains an explicit reliability option for unattended machines. It is not part of the default installation.

There is no universal API that can force every possible LLM website to access local files. “Any local AI” therefore means any agent application that can read files, run a command, and retain a persistent instruction. The engine itself contains no model- or provider-specific adapter.

Karpathy's LLM Wiki idea is powerful because knowledge is compiled and maintained once instead of rediscovered for every question. Enhanced Second Brain adds a portable OKF contract, fast local retrieval, usage-aware maintenance, and reversible cleanup while preserving that simple model.

SQLite FTS5 is deliberately the default because it is mature, embedded, cross-platform, and zero-configuration. The technical overview explains the database choice and when a replicated alternative would make sense.

The ingredients are not unprecedented. Other projects already provide Markdown personal wikis, semantic second brains, or persistent agent memory. The contribution here is a deliberately small, provider-neutral combination focused on portable OKF files, zero-model lexical retrieval, usage-aware reversible cleanup, and one-file export. See what is actually distinctive.

The normal context path is also optimized for repeated agent requests: it uses one process, one SQLite connection, one locked telemetry append, and skips excluded directories during freshness checks. The release uses a self-contained directory bundle rather than a compressed one-file executable, because repeatedly unpacking that executable was much slower. See the runtime benchmark and exact scope.

• A local agent needs filesystem and command access. A browser-only chat cannot silently read local files.
• Lexical search can miss paraphrases that share no useful vocabulary; benchmark your own corpus.
• Obsidian is optional and remains only an interface for the same portable files.
• Remote backup is opt-in and must point to a private destination.
• A portable ZIP is not encrypted automatically; protect it like any private copy of your memory.

Apache-2.0 licensed.