sandclaw-memory

Self-Growing Tag-Dictionary RAG for Any Device

---

Give your AI long-term memory that grows smarter over time.

No GPU. No vector database. No external dependencies. Just pip install and go.

from sandclaw_memory import BrainMemory

brain = BrainMemory(tag_extractor=my_ai_func)
brain.save("User loves Python and React")
context = brain.recall("what does the user like?")
# -> Returns relevant memories as Markdown, ready for LLM injection

Why sandclaw-memory?

Feature	sandclaw-memory	Vector DB (Pinecone, Weaviate)	mem0
Setup	pip install (done)	Docker + API keys + config	pip install + API key
GPU required	No	Often yes	No
Cost over time	Decreases (self-growing)	Constant	Constant
Search	FTS5 + BM25 (proven)	Cosine similarity	Embedding-based
Privacy	100% local	Cloud required	Cloud optional
Dependencies	0 (stdlib only)	Many	Several
Size	~50KB	100MB+ Docker	~5MB

The Self-Growing Advantage

Traditional RAG calls AI for every search. sandclaw-memory learns:

Day 1:  "React로 페이지 만듦"  → AI extracts: [react, frontend, page]
         keyword_map registers: "React" → react, "페이지" → page

Day 30: "React Native 시작"   → keyword_map instant match! No AI needed.
         Cost: $0.00

Day 90: 80% of saves match keyword_map → AI calls reduced by 80%

The more you use it, the cheaper it gets.

Quick Start

Install

pip install sandclaw-memory

3-Line Memory

from sandclaw_memory import BrainMemory

with BrainMemory(tag_extractor=my_tag_func) as brain:
    brain.save("Important: migrate to FastAPI by Q2")
    print(brain.recall("what's the migration plan?"))

With OpenAI

import json, openai
from sandclaw_memory import BrainMemory

def tag_extractor(content: str) -> list[str]:
    resp = openai.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{
            "role": "system",
            "content": "Extract 3-7 keyword tags. Return JSON array only."
        }, {"role": "user", "content": content}],
        temperature=0,
    )
    return json.loads(resp.choices[0].message.content)

with BrainMemory(
    db_path="./my_memory",
    tag_extractor=tag_extractor,
) as brain:
    brain.start_polling()  # Background tag extraction every 15s

    brain.save("User prefers Python and TypeScript")
    brain.save("Decided to use PostgreSQL", source="archive")

    context = brain.recall("what tech stack?")
    # Inject 'context' into your LLM's system prompt

With Claude

import json, anthropic
from sandclaw_memory import BrainMemory

client = anthropic.Anthropic()

def tag_extractor(content: str) -> list[str]:
    resp = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=200,
        messages=[{"role": "user",
                   "content": f"Extract 3-7 tags as JSON array:\n{content}"}],
    )
    return json.loads(resp.content[0].text)

brain = BrainMemory(tag_extractor=tag_extractor)

See examples/ for LangChain, custom callbacks, and scheduling patterns.

Architecture

┌─────────────────────────────────────────────────┐
│                  BrainMemory                     │
│    save() → recall() → start_polling()          │
├─────────────────────────────────────────────────┤
│                                                  │
│  ┌─────────┐  ┌──────────┐  ┌────────────────┐  │
│  │   L1    │  │    L2    │  │      L3        │  │
│  │ Session │  │ Summary  │  │   Archive      │  │
│  │         │  │          │  │                │  │
│  │ 3-day   │  │ 30-day   │  │ Permanent      │  │
│  │ rolling │  │ AI/text  │  │ SQLite + FTS5  │  │
│  │ Markdown│  │ summary  │  │ + self-growing │  │
│  │  logs   │  │          │  │   tag dict     │  │
│  └────┬────┘  └────┬─────┘  └───────┬────────┘  │
│       │            │                │            │
│  ┌────┴────────────┴────────────────┴────────┐  │
│  │          Intent Dispatcher                 │  │
│  │  CASUAL → L1  │ STANDARD → L1+L2         │  │
│  │               │ DEEP → L1+L2+L3           │  │
│  └───────────────┴───────────────────────────┘  │
│                                                  │
│  ┌──────────────────────────────────────────┐   │
│  │     Self-Growing Tag Pipeline            │   │
│  │  Stage 1: keyword_map (instant, free)    │   │
│  │  Stage 2: AI callback (async, queue)     │   │
│  │  → keyword_map grows → Stage 2 shrinks   │   │
│  └──────────────────────────────────────────┘   │
└─────────────────────────────────────────────────┘

Three Memory Layers

Layer	What	Retention	Storage	Speed
L1 Session	Conversation logs	3 days (rolling)	Markdown files	Instant
L2 Summary	Period summaries	30 days	In-memory	Instant
L3 Archive	Important memories	Forever	SQLite + FTS5	~1ms

Intent-Based Depth

The dispatcher auto-detects how deep to search:

brain.recall("what time is it?")           # → CASUAL  (L1 only)
brain.recall("summarize this month")       # → STANDARD (L1 + L2)
brain.recall("why did we pick React?")     # → DEEP    (L1 + L2 + L3)

# Or override manually:
brain.recall("anything", depth="deep")

API Reference

BrainMemory

BrainMemory(
    db_path="./memory",           # Where to store files
    tag_extractor=my_func,        # REQUIRED: (str) -> list[str]
    promote_checker=None,         # Optional: (str) -> bool
    depth_detector=None,          # Optional: (str) -> str
    duplicate_checker=None,       # Optional: (str, str) -> bool
    conflict_resolver=None,       # Optional: (str, str) -> str
    polling_interval=15,          # Seconds between maintenance cycles
    rolling_days=3,               # L1 retention period
    summary_days=30,              # L2 summary period
    max_context_chars=15_000,     # Context budget for LLM injection
    encryption_key=None,          # Optional SQLCipher encryption
)

Core Methods

Method	Description
save(content, source="chat", tags=None)	Save to memory. source="archive" forces L3.
recall(query, depth=None)	Recall relevant memories as Markdown.
search(query, limit=20)	Search L3 archive, returns list[MemoryEntry].
promote(content, tags=None)	Manually save to L3. Returns memory ID.
summarize(llm_callback=None)	Generate a 30-day summary.
start_polling()	Start background maintenance loop.
stop_polling()	Stop background maintenance loop.
run_maintenance()	Run one maintenance cycle manually.
get_stats()	Get stats from all layers.
get_tag_stats()	Get tag usage counts.
export_json(path=None)	Export all data as JSON.
on(event, callback)	Register an event hook.
close()	Stop polling and close DB.

The 5 AI Callbacks

Callback	Signature	Default
tag_extractor	(str) -> list[str]	REQUIRED
promote_checker	(str) -> bool	len(content) > 200
depth_detector	(str) -> str	Keyword matching
duplicate_checker	(str, str) -> bool	SequenceMatcher > 0.85
conflict_resolver	(str, str) -> str	Keep newer text

Event Hooks

brain.on("before_save",    lambda content, source: ...)
brain.on("after_save",     lambda content, source: ...)
brain.on("before_promote", lambda content: ...)
brain.on("after_promote",  lambda content, mem_id: ...)
brain.on("after_recall",   lambda query, depth, result: ...)
brain.on("after_cycle",    lambda stats: ...)

Examples

File	Description
basic_usage.py	Simplest setup with OpenAI
with_openai.py	All 5 callbacks with OpenAI
with_anthropic.py	Claude API integration
with_langchain.py	LangChain agent + memory
custom_callbacks.py	Custom callbacks + hooks
scheduling.py	Polling, FastAPI, Celery

Compatibility

• Python: 3.9, 3.10, 3.11, 3.12, 3.13
• OS: Windows, macOS, Linux
• Dependencies: None (Python stdlib only)
• SQLite: Uses built-in sqlite3 module (FTS5 optional, graceful fallback)

Roadmap

Version	Codename	Status
v0.1.0	"Remembering AI"	Current
v0.2.0	"Growing AI"	Planned -- tag trees + built-in AI (Gemma 4)
v0.3.0	"Thinking AI"	Planned -- contradiction detection (CKN)

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT License. See LICENSE for details.