Self-Learning Agent

Use a three-layer memory system instead of one giant memory file.

Goal

Keep session startup cheap while preserving durable knowledge.

Architecture

workspace/
├── MEMORY.md
└── memory/
    ├── cards/
    │   ├── topic-name.md
    │   └── another-topic.md
    └── YYYY-MM-DD.md

Model

• MEMORY.md: slim index loaded every session
• memory/cards/*.md: curated knowledge cards, one topic per file
• memory/YYYY-MM-DD.md: raw daily session logs

Why this works

• The index stays small.
• Durable knowledge is split into searchable units.
• Raw logs do not pollute the always-loaded context.
• The agent loads only the memory it needs for the current task.

Session workflow

1. Read MEMORY.md.
2. Search for task-relevant cards.
3. Skim recent daily logs when recency matters.
4. Work.
5. Write back durable lessons as cards.
6. Log the session briefly in the daily file.

Card format

---
topic: Descriptive Topic Name
category: system|human|infrastructure|tools|workflow|projects|lessons|security|models
tags: [tag1, tag2, tag3]
created: YYYY-MM-DD
updated: YYYY-MM-DD
---

Dense, factual content.
Include specific commands, decisions, paths, or gotchas when they matter.
Keep it short enough to stay easy to retrieve and update.

Card rules

1. One topic per card.
2. Prefer updates over duplicates.
3. Write for future-you with zero context.
4. Store decisions and lessons, not chatter.
5. Keep examples concrete.

Capture triggers

Write or update a card when you learn:

• a hard-won fix
• a recurring workflow
• a user correction
• a non-obvious system fact
• a durable preference or rule

Do not capture

• trivial facts
• one-off noise
• full conversation transcripts
• stale information you already replaced elsewhere

Daily log format

## HH:MM - Brief title

What happened, what changed, what was learned.
Link any new or updated cards.

Maintenance loop

Periodically:

1. review recent daily logs
2. promote durable lessons into cards
3. merge duplicate cards
4. prune outdated index entries

Anti-pattern

Do not keep appending everything to one massive memory file. That turns recall into sludge and makes every future session more expensive.