self-learning-agent（自学习 Agent）

Self-Learning Agent — Knowledge Card Memory System

A production-tested memory architecture for AI agents that wake up fresh each session. Instead of one monolithic memory file that grows until it's unusable, this system uses atomic knowledge cards (~350 tokens each) searched semantically, daily logs for raw notes, and a slim master index loaded every session.

Architecture

workspace/
├── MEMORY.md              # Master index (~2KB, loaded every session)
├── memory/
│   ├── cards/             # Knowledge cards (~350 tokens each)
│   │   ├── topic-name.md  # One topic per file, YAML frontmatter
│   │   ├── another-topic.md
│   │   └── ...
│   └── YYYY-MM-DD.md      # Daily session logs (raw notes)

Why This Works

• MEMORY.md is tiny (~2KB). It loads fast, gives the agent orientation, and points to everything else.
• Knowledge cards are atomic. Each one covers ONE topic in ~350 tokens. Semantic search finds the right cards without loading everything.
• Daily logs are append-only scratch pads. Raw session notes, not curated.
• Cards are curated wisdom. Daily logs are raw data. The agent periodically distills daily logs into cards during maintenance.

Setup

1. Create the directory structure

mkdir -p memory/cards

2. Create MEMORY.md (master index)

This file is loaded every session. Keep it under 2KB. It should contain:

# MEMORY.md — Master Index

## How Memory Works
- **This file:** Slim index (~2KB). Loaded every main session.
- **Knowledge cards:** `memory/cards/*.md` (~N cards, ~350 tokens each). Searched semantically.
- **Daily logs:** `memory/YYYY-MM-DD.md`. Raw session notes.
- **DO NOT** dump everything here. Write knowledge cards instead.

## Identity
[Agent name, model, owner, key facts]

## Quick Context
[2-3 lines of what matters right now]

## Card Categories
[Table mapping categories to card topics]

## Current Priorities
[What's actively being worked on]

3. Add to your AGENTS.md / system prompt

## Every Session
1. Read MEMORY.md (slim index)
2. Search `memory_search` for context relevant to the current task
3. Skim today + yesterday daily logs for recent context
4. Start working

## Memory Rules
- "Mental notes" don't survive session restarts. Files do.
- When someone says "remember this" → write a knowledge card
- When you learn a lesson → write a knowledge card
- When you make a mistake → document it so future-you doesn't repeat it

Knowledge Card Format

Every card has YAML frontmatter and dense content:

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

The actual content. Dense, factual, no fluff.
Write for future-you who has zero context.
Include specific commands, paths, config values.
Keep under 350 tokens.

Card Quality Rules

1. ONE topic per card. Three insights = three cards.
2. ~350 tokens max. Dense beats verbose.
3. Zero-context readable. Include specifics (commands, paths, values).
4. Tags are searchable keywords. Lowercase, hyphenated.
5. Update, don't duplicate. If a card exists for the topic, merge new info into it.
6. No fluff. Every sentence should contain a fact, a command, or a decision.

Good Card Example

---
topic: Cortex CSRF Automation
category: infrastructure
tags: [cortex, csrf, thehive, api, security]
created: 2026-03-19
updated: 2026-03-19
---

Cortex 3.1.8 uses non-standard CSRF. Cookie: CORTEX-XSRF-TOKEN, header: X-CORTEX-XSRF-TOKEN.
Standard Play Framework bypass headers (Csrf-Token: nocheck) do NOT work.

Flow: Login → GET any endpoint with session cookie → capture CORTEX-XSRF-TOKEN from Set-Cookie →
send as both cookie AND X-CORTEX-XSRF-TOKEN header on all POST/PUT/DELETE.

Shortcut: After generating first API key, use Authorization: Bearer which bypasses CSRF entirely.
First-user POST /api/user (no auth) only works when zero users exist in DB.

Bad Card Example

---
topic: Stuff I Learned Today
---

Today I learned a bunch of things about Cortex and TheHive. The CSRF thing was really tricky
and took a while to figure out. I also learned about how to set up organizations and users.
It was a productive session overall.

(Too vague, no specifics, no actionable info, multiple topics in one card)

Capture Triggers

Automatic (agent should capture without being asked)

• Hard-won debugging lessons (3+ attempts to fix something)
• Configuration gotchas (things that work differently than expected)
• User corrections ("no, do it THIS way")
• Non-obvious facts about infrastructure, people, or projects
• Workflow improvements discovered during a task

Manual

• User says /learn, "remember this", or "save this"
• User explicitly corrects the agent's approach

What NOT to Capture

• Obvious/trivial information
• Temporary context (one-time fixes that won't recur)
• Things already in existing cards
• Conversation summaries (that's what daily logs are for)

Daily Log Format

Append to memory/YYYY-MM-DD.md:

## HH:MM — Brief Title

What happened, what was decided, what was learned.
Link to any cards created: `→ card: topic-name`

Memory Maintenance

Periodically (every few days), the agent should:

1. Read recent daily logs
2. Identify significant events worth preserving long-term
3. Create or update knowledge cards from insights
4. Remove outdated info from MEMORY.md
5. Update the card categories table in MEMORY.md

Think of it like a human reviewing their journal and updating their mental model.

Promotion Rules

When the same lesson appears 3+ times in cards:

• Promote it to AGENTS.md as a permanent rule
• Mark the original card as "promoted"
• This prevents the agent from re-learning the same lesson

Session Workflow

Session Start
    │
    ├── Read MEMORY.md (always, ~2KB)
    ├── memory_search for task-relevant cards
    ├── Skim today + yesterday daily logs
    │
    ├── [Do work]
    │
    ├── Capture insights → knowledge cards
    ├── Log session → daily log
    │
Session End

Scaling

This system has been tested with:

• ~36 knowledge cards (~350 tokens each = ~12.6K tokens total)
• Daily logs spanning months
• Semantic search via embeddings (qwen3-embedding or similar)

At this scale, semantic search finds relevant cards in <100ms. The master index stays under 2KB. The agent loads only what it needs.

If you hit 100+ cards, consider:

• Archiving cards older than 6 months that haven't been accessed
• Splitting categories into subdirectories
• Adding a card index file per category

Comparison with Monolithic Memory