Skip to content

README_EN.md

Latest commit

ย 

History

History
321 lines (234 loc) ยท 7.84 KB

README_EN.md

File metadata and controls

321 lines (234 loc) ยท 7.84 KB

๐Ÿงฌ DNA Memory

Make AI Agents learn, reinforce, forget, and generalize like a human brain

Stars License Python OpenClaw

English | ็ฎ€ไฝ“ไธญๆ–‡

Most AI memory systems only solve storage. DNA Memory is about how agents actually learn and evolve.

It is not just a memory store. It is a memory evolution system with:

  • 3-layer memory architecture
  • reinforcement and decay
  • reflection (reflect)
  • promotion to long-term memory (promote)
  • duplicate cleanup (dedupe)
  • FTS5-powered recall search
  • background daemon maintenance

๐Ÿ†š Why not just use a normal memory store?

Capability Mem0 Zep LangChain Memory DNA Memory
Basic storage โœ… โœ… โœ… โœ…
Vector / semantic retrieval โœ… โœ… โœ… โš ๏ธ extensible
Multi-layer architecture โŒ โš ๏ธ โŒ โœ… working / short / long
Active forgetting โŒ โŒ โŒ โœ…
Reflection loop โŒ โŒ โŒ โœ…
Pattern extraction โŒ โŒ โŒ โœ…
Long-term promotion โŒ โŒ โŒ โœ…
Local-first / minimal core deps โŒ โŒ โŒ โœ…
Built for agent workflows โš ๏ธ โš ๏ธ โš ๏ธ โœ…

Positioning in one sentence:

DNA Memory helps AI agents not only remember, but also reinforce, forget, summarize, and evolve like a real cognitive system.

๐Ÿš€ Quick Start in 30 Seconds

# 1) Clone into your OpenClaw skills directory
git clone https://github.com/AIPMAndy/dna-memory.git ~/.openclaw/skills/dna-memory

# 2) Remember one preference
python3 ~/.openclaw/skills/dna-memory/scripts/evolve.py remember "The user prefers concise and direct responses" -t preference -i 0.9

# 3) Recall related memories
python3 ~/.openclaw/skills/dna-memory/scripts/evolve.py recall "concise direct"

# 4) Inspect stats
python3 ~/.openclaw/skills/dna-memory/scripts/evolve.py stats

Why it is practical:

  • core features run on Python + SQLite
  • no external database required
  • local-first by default
  • ideal for personal assistants, local agents, and autonomous workflows

โœจ Core Capabilities

1. Three-layer memory architecture

Working Memory
  โ†“ filter
Short-term Memory
  โ†“ consolidate / promote
Long-term Memory
Layer Role Typical content
Working temporary session context current task state, fresh facts
Short-term recent important information preferences, lessons, recent errors
Long-term stable knowledge and patterns rules, skills, persistent preferences

2. Reinforcement and forgetting

  • used often โ†’ higher weight
  • unused for a long time โ†’ decay
  • low-weight memories โ†’ removable
  • stable high-value memories โ†’ promoted to long-term memory

3. Reflection (reflect)

reflect does two things:

  • extracts recurring patterns from recent high-weight memories
  • promotes stable short-term memories into long-term memory

4. Better recall search

Current recall supports:

  • multi-keyword AND search
  • type filters like type:error / type:skill
  • SQLite FTS5 full-text search
  • automatic fallback to LIKE search if FTS5 is unavailable

Examples:

python3 scripts/evolve.py recall "feishu api"
python3 scripts/evolve.py recall "type:error github"
python3 scripts/evolve.py recall "user preference concise"

5. Background maintenance daemon

The daemon can automatically run:

  • reflect
  • decay
  • throttled maintenance so the same batch is not repeatedly summarized

It can also be registered with macOS launchd for auto-start on boot.

๐Ÿ“ฆ Actual current architecture

dna-memory/
โ”œโ”€โ”€ scripts/
โ”‚   โ”œโ”€โ”€ evolve.py              # core CLI: remember / recall / stats / reflect / dedupe ...
โ”‚   โ”œโ”€โ”€ dna_memory_daemon.py   # background maintenance daemon
โ”‚   โ”œโ”€โ”€ semantic_search.py     # experimental semantic search module
โ”‚   โ”œโ”€โ”€ analyze.py
โ”‚   โ”œโ”€โ”€ api.py
โ”‚   โ”œโ”€โ”€ autocollect.py
โ”‚   โ”œโ”€โ”€ backup.py
โ”‚   โ”œโ”€โ”€ cli.py
โ”‚   โ”œโ”€โ”€ detailed_stats.py
โ”‚   โ”œโ”€โ”€ knowme_link.py
โ”‚   โ”œโ”€โ”€ reminder.py
โ”‚   โ”œโ”€โ”€ trigger.py
โ”‚   โ””โ”€โ”€ visualize.py
โ”œโ”€โ”€ memory/
โ”‚   โ”œโ”€โ”€ memory.db              # SQLite primary store (memories + operations)
โ”‚   โ””โ”€โ”€ working.json           # working memory
โ”œโ”€โ”€ assets/
โ”‚   โ””โ”€โ”€ config.json            # daemon / decay config
โ”œโ”€โ”€ README.md
โ”œโ”€โ”€ README_EN.md
โ””โ”€โ”€ SKILL.md

Note: memory/*.db should not be committed. The repo now ignores real memory database files by default.

๐Ÿงช Core Commands

Remember

python3 scripts/evolve.py remember "Andy prefers concise and direct responses" -t preference -i 0.95

Recall

python3 scripts/evolve.py recall "concise response"
python3 scripts/evolve.py recall "type:skill feishu"

Stats

python3 scripts/evolve.py stats

Reflect

python3 scripts/evolve.py reflect

Promote

python3 scripts/evolve.py promote --id 12

Dedupe

python3 scripts/evolve.py dedupe

Daemon

# start
python3 scripts/dna_memory_daemon.py start

# check status
python3 scripts/dna_memory_daemon.py status

# stop
python3 scripts/dna_memory_daemon.py stop

โš™๏ธ Use Cases

1. Personal AI assistants

  • remember user preferences
  • develop a stable collaboration style over time
  • learn from mistakes instead of repeating them

2. Agent workflow orchestration

  • turn finished tasks into reusable skills
  • store failure cases as error memories
  • extract patterns from long-running work

3. AI products with personalization

  • accumulate user profiles
  • track behavioral patterns
  • build long-term personalization

4. Self-improving agent systems

  • works well with OpenClaw, self-improving-agent, and custom agent stacks
  • turns operational experience into reusable memory assets

๐Ÿงญ Recommended Workflow

Receive task
  โ†“
Recall related memories
  โ†“
Execute
  โ†“
Remember new preferences / skills / errors
  โ†“
Reflect recurring patterns
  โ†“
Promote into long-term memory

This workflow is especially useful when:

  • the user corrects the agent
  • a new preference is learned
  • an API/tool fails
  • a long task finishes
  • a repeatable pattern appears

๐Ÿ—บ๏ธ Roadmap

  • SQLite single-store refactor
  • remember / recall / reflect / promote / dedupe CLI
  • daemon for automatic reflect / decay
  • FTS5-based recall search
  • launchd auto-start setup
  • better Chinese tokenization and ranking
  • real embedding-based semantic retrieval
  • stronger memory graph visualization
  • more complete import / export / migration tooling
  • shared memory spaces for multi-agent systems

๐Ÿค Contributing

Issues and PRs are welcome.

High-impact contribution areas:

  • recall ranking quality
  • Chinese search experience
  • pattern extraction quality
  • memory visualization
  • embedding provider integrations

๐Ÿ‘จโ€๐Ÿ’ป Author

Andy / AI้…‹้•ฟAndy
Ex-Tencent / Baidu AI Product Expert โ†’ LLM Unicorn VP โ†’ Startup CEO

Focus areas:

  • AI agents
  • AI commercialization
  • memory systems
  • human augmentation

GitHub: https://github.com/AIPMAndy

๐Ÿ“„ License

Apache 2.0

If this project helps you, give it a โญ Star.