Advisor Memory

A bilingual 中英混合 advisor-style memory skill for Codex and Claude-style agents.

advisor-memory turns external corpora, project notes, and durable judgment pages into one shared knowledge base for both humans and agents. It uses GBrain as the durable memory substrate, but keeps the source library human-readable and usable even without embeddings.

What It Is

This is not just a search wrapper around a document folder.

It is a memory system with three jobs:

• preserve raw source material
• route questions across multiple registered corpora
• accumulate durable project and concept memory over time

It is designed for cases like:

• “Lenny 怎么说 onboarding / PMF / pricing?”
• “把这批访谈资料吸收到记忆里”
• “结合我当前项目，给我一个带参考依据的建议”
• “跨多个库一起搜，不要让我先选 pack”

First Run After Install

If you just cloned this skill into ~/.claude/skills/advisor-memory or ~/.codex/skills/advisor-memory, the first thing to do is bootstrap it.

Recommended first command:

python3 scripts/bootstrap_advisor_memory.py \
  --brain-root ~/brain \
  --lenny-root /absolute/path/to/lennys-podcast-newsletter \
  --install-gbrain

This one command will:

• install or locate gbrain
• run gbrain init --pglite
• run gbrain doctor --json
• create the advisor-memory brain layout
• register Lenny as the first pack
• write ~/.advisor-memory/config.json
• print the first commands you should run next

If you do not want to start with Lenny yet, you can omit --lenny-root and bootstrap an empty brain first:

python3 scripts/bootstrap_advisor_memory.py --brain-root ~/brain --install-gbrain

Product Model

The core design rule is simple:

• one canonical knowledge base
• human-readable by default
• agent-optimized routing on top

So this system does not create one memory base for humans and another for agents.

Instead, it keeps:

• human-facing entrypoints in normal markdown pages and pack index.md
• machine-facing metadata under _system/
• raw files preserved in raw/

Core Capabilities

• Layered memory: working/ for short-term context, durable pages for long-term memory, sources/libraries/ for registered corpora.
• Human-readable packs: every pack exposes an index.md; metadata and routing artifacts live under _system/.
• Broad advisory retrieval by default: users should not need to choose a pack first.
• Project-aware consults: if a known project is mentioned, the project page becomes the primary frame and external libraries become evidence.
• Raw-first reasoning: summaries are used for routing; full raw documents are preferred for real reasoning whenever possible.
• Automatic pack registration: new corpora can be ingested and indexed without hand-building the structure.
• Cross-pack search: the system can route across multiple packs instead of treating each corpus as a separate workflow.
• Dream cycle maintenance: stale summaries, maps, and promotion queues can be refreshed automatically.
• Embedding-optional day one: without an OpenAI embedding key, the system still works in keyword/fulltext/raw mode.

Directory Model

brain/
├── people/
├── companies/
├── projects/
├── concepts/
├── ideas/
├── working/
│   ├── active-topics/
│   ├── recent-conversations/
│   ├── pending-promotions/
│   └── project-memory-reviews/
├── sources/
│   ├── domain_taxonomy.json
│   ├── domain_taxonomy.generated.json
│   └── libraries/
│       ├── index.md
│       └── <pack>/
│           ├── index.md
│           ├── raw/
│           └── _system/
│               ├── manifest.json
│               ├── indexes/
│               ├── summaries/
│               └── domain_suggestions.json
└── inbox/

Human-facing knowledge usually starts from:

• projects/, concepts/, people/, companies/, ideas/
• sources/libraries/index.md
• each pack’s index.md
• raw/ when deeper reading is needed

How It Works

1. Ingest

You give the system a file or a corpus.

It preserves the raw source, registers it as a library pack, builds indexes and summaries, creates a human-facing index.md, and keeps machine metadata under _system/.

2. Consult

You ask a question.

The default behavior is:

1. read durable memory first
2. detect relevant domains
3. route across the most relevant packs
4. drill into the best raw sources
5. produce an advisory answer instead of a corpus summary

If the question mentions a known project, the project page becomes the primary context.

3. Write Back

If a consult produces a durable insight, the system can:

• emit a candidate project-memory review
• ask for confirmation
• write confirmed updates back into projects/...

4. Consolidate

During a dream cycle, the system can:

• refresh stale routing artifacts
• emit promotion queues
• refresh GBrain

Quick Start

Prerequisites

• Python 3
• gbrain
• a local corpus to ingest, for example the Lenny archive

Install gbrain:

bun add -g github:garrytan/gbrain
gbrain --version

Initialize a Lenny-first brain

Preferred one-command path:

python3 scripts/bootstrap_advisor_memory.py \
  --brain-root /absolute/path/to/brain \
  --lenny-root /absolute/path/to/lennys-podcast-newsletter \
  --install-gbrain

Low-level manual path:

python3 scripts/init_lenny_first_memory.py \
  --brain-root /absolute/path/to/brain \
  --lenny-root /absolute/path/to/lennys-podcast-newsletter

Ask a broad advisory question

python3 scripts/consult_advisor_memory.py \
  --brain-root /absolute/path/to/brain \
  "How should we improve onboarding and PMF?"

Ingest a new corpus

python3 scripts/smart_ingest_library_pack.py \
  --brain-root /absolute/path/to/brain \
  --source /absolute/path/to/new-corpus

Search across all packs

python3 scripts/search_library_pack.py \
  --brain-root /absolute/path/to/brain \
  --all-packs search "onboarding product-market fit" --limit 10

Run a dream cycle

python3 scripts/dream_cycle.py --brain-root /absolute/path/to/brain --apply

No Embedding Required

This repo is designed to be useful on day one without an OpenAI embedding key.

Without embeddings, it still supports:

• pack registration
• raw preservation
• search / fulltext / read
• cross-pack consult
• GBrain keyword retrieval

If you later add an OpenAI key, gbrain embed --stale enables stronger semantic retrieval on top of the same workflow.

Project Memory

Project pages are intended to grow naturally.

The system does not force a rigid project dashboard template. Instead:

• projects/project-a.md acts as the durable project memory page
• consults can generate candidate updates
• confirmed updates can be written back
• external libraries serve as evidence, not as the primary project memory

This keeps project memory useful without over-structuring it.

Human Use vs Agent Use

This repository deliberately keeps the knowledge base usable for both:

• Humans can browse markdown pages, pack indexes, and raw files directly.
• Agents can use _system/manifest.json, _system/indexes/, and _system/summaries/ for routing and automation.

No separate viewer product is required for the system to be usable by humans.

Repository Contents

• SKILL.md — the skill contract and operating rules
• agents/openai.yaml — skill metadata for OpenAI/Codex-style environments
• references/ — architecture, init, pack protocol, workflows
• scripts/ — deterministic runtime and maintenance scripts

Recommended Entry Points

• First-run bootstrap: scripts/bootstrap_advisor_memory.py
• Setup: scripts/init_lenny_first_memory.py
• Ingest: scripts/smart_ingest_library_pack.py
• Broad consult: scripts/consult_advisor_memory.py
• Low-level pack search: scripts/search_library_pack.py
• Consolidation: scripts/dream_cycle.py
• Durable promotion: scripts/promote.py
• Project-memory writeback: scripts/update_project_memory.py

Status

advisor-memory is a production-usable skill focused on:

• long/short-term layered memory
• human-readable source libraries
• broad advisory retrieval
• project-aware recommendations

Current scope is intentionally narrow:

• no standalone UI product
• no built-in transcript/audio transcription layer
• domain extension is heuristic-first rather than fully semantic

Read More

• SKILL.md
• references/architecture.md
• references/init.md
• references/library-packs.md
• references/workflows.md

许可证

MIT License,见 LICENSE。