comet-memory 0.4.0

CoMeT — Cognitive Memory Tree: Hierarchical memory system for LLM agents

☄️ CoMeT — Cognitive Memory Tree

Lossless structured memory for AI agents.

Recent Updates

• 🚀 3-Tier Progressive Retrieval: Short summary → Lazy detailed summary → Raw content
• 🔗 GCRI Integration: In-session memory for multi-agent reasoning with auto-ingest
• 📄 Document Ingestion: add_document() for chunked ingestion of large texts

CoMeT compresses long conversations into a navigable tree of memory nodes.
Unlike naive summarization that loses details, CoMeT preserves full raw content behind structured summaries — agents read summaries first, then progressively drill deeper only when needed.

Architecture

User Input / Document
     │
     ▼
┌─────────┐    SLM (fast)     ┌───────────┐
│  Sensor  │ ───────────────▶ │ L1 Buffer │
└─────────┘   entity/intent   └─────┬─────┘
                                    │ cognitive load trigger
                                    ▼
                              ┌───────────┐
                              │ Compacter │  LLM (slow)
                              └─────┬─────┘
                                    │ summary + trigger + recall_mode + tags
                                    ▼
                         ┌──────────┴──────────┐
                         │                     │
                   ┌───────────┐        ┌─────────────┐
                   │   Store   │        │ VectorIndex │  ChromaDB
                   │  depth 0-2│        │  full raw   │  summary + trigger
                   └───────────┘        └──────┬──────┘
                                               │
                                    ┌──────────┼──────────┐
                                    ▼          ▼          ▼
                              ┌───────────────────────────────┐
                              │         3-Tier Retrieval       │
                              │  T1: Summary  (always cached)  │
                              │  T2: Detailed  (lazy, on-demand)│
                              │  T3: Raw       (full original)  │
                              └───────────────────────────────┘

Dual-Speed Layer

• Fast (Sensor): SLM extracts entities/intent per turn, detects topic shifts via cognitive load assessment
• Slow (Compacter): Main LLM structures accumulated L1 buffer into MemoryNode with summary, trigger, recall mode, and topic tags

3-Tier Progressive Retrieval

Agents retrieve information at increasing depth, paying token cost only when needed:

Tier	Method	Content	Token Cost
1	retrieve	Short summary + trigger + node_id	Minimal
2	get_detailed_summary	3–8 sentence detailed summary	Medium (lazy-generated, then cached)
3	get_raw_content	Full original content	Full

Lazy Detailed Summary: Tier 2 summaries are generated on first request from raw content via SLM, then cached in the node. Subsequent calls return the cached version at zero additional cost.

Recall Mode

Each memory node is classified by recall_mode at compaction time:

Mode	Behavior	Examples
passive	Always included in context window	User identity, persistent preferences
active	Retrieved on-demand via semantic search	Factual details, decisions, events
both	Always in context + searchable via RAG	Core constraints with retrievable details

Dual-Path RAG Retrieval

CoMeT embeds both summary (what the node contains) and trigger (when to recall it) into separate vector collections. At query time:

1. QueryAnalyzer decomposes the query into semantic_query + search_intent
2. Summary path: matches what the information is about
3. Trigger path: matches when the information would be needed
4. ScoreFusion (Reciprocal Rank Fusion): merges results from both paths

Triggers are written from the LLM's perspective ("내가 ~정보가 필요할 때") rather than user-centric, enabling broader semantic matching even without explicit user requests.

Document Ingestion

Large documents and tool outputs can be ingested directly via add_document():

nodes = memo.add_document(
    content=long_text,
    source='tool:search_web',
    chunk_size=2000,
    chunk_overlap=200
)

Text is split into overlapping chunks at sentence/line boundaries, each processed through the Sensor → Compacter pipeline. Full raw content is stored in the vector store without truncation.

Consolidation

Cross-session deduplication, linking, and tag normalization:

1. Dedup: Detect and merge semantically similar nodes
2. Cross-link: Create bidirectional links between related (non-duplicate) nodes
3. Tag normalization: Unify variant tags that refer to the same concept

Topic-Aware Auto-Linking

Nodes share a global topic tag set. The compacter reuses existing tags when possible, enabling automatic bidirectional linking between related nodes across different conversation segments.

Benchmark (52 turns, 5 conversations, 10 questions)

Method	Context Cost	Accuracy
Full Context Injection	5,198 chars (100%)	10/10
CoMeT	1,397 chars (27%)	9/10
Naive Summary	1,179 chars (23%)	1/10

• CoMeT uses 27% of the tokens while retaining 90% accuracy
• 6/10 questions required link traversal (agent read 2-3 nodes)
• Cross-topic questions: CoMeT 5/5 vs Naive 0/5

Quick Start

Session Memory (within a conversation)

from comet import CoMeT, scope

@scope
def main(config):
    memo = CoMeT(config)

    # Add conversation turns
    memo.add("B200 4대로 월드모델 학습 가능할까?")
    memo.add("2B면 충분하고 커봐야 8B")
    memo.add("DPO 데이터는 negative를 syntax error로 구성했어")

    # Force compact remaining buffer
    memo.force_compact()

    # Navigation
    for node in memo.list_memories():
        print(memo.read_memory(node['node_id'], depth=0))

    # Agent tools (LangChain compatible)
    tools = memo.get_tools()
    # → get_memory_index, read_memory_node, search_memory

main()

Cross-Session RAG Retrieval

from comet import CoMeT, scope

@scope
def main(config):
    config.retrieval.vector_db_path = './memory_store/vectors'

    memo = CoMeT(config)

    # Ingest turns (auto-indexed to VectorIndex on compaction)
    memo.add("JWT 액세스 토큰 만료는 15분, 리프레시는 7일로 설정")
    memo.force_compact()

    # Semantic retrieval across all sessions
    results = memo.retrieve("토큰 만료 설정이 어떻게 되어있어?")
    for r in results:
        print(f"[{r.node.node_id}] score={r.relevance_score:.4f}")
        print(f"  {r.node.summary}")

main()

3-Tier Progressive Retrieval

# Tier 1: Short summary scan
results = memo.retrieve("LangGraph architecture")
# → [mem_xxx] (score=0.85) LangGraph 프레임워크 아키텍처 요약

# Tier 2: Lazy detailed summary (generated on first call, cached after)
detailed = memo.get_detailed_summary("mem_xxx")
# → "LangGraph provides graph-based orchestration with checkpointing..."

# Tier 3: Full raw content (only when needed)
raw = memo.get_raw_content("mem_xxx")
# → [complete original text]

Document Ingestion

# Ingest large documents (auto-chunked)
nodes = memo.add_document(
    content=web_search_result,
    source='https://example.com/article'
)

GCRI Integration

CoMeT serves as the in-session memory layer for GCRI (Graph-based Collective Reasoning Intelligence), a multi-agent reasoning framework.

3-Tier Tool Pipeline

GCRI agents access CoMeT through three progressively deeper tools:

Tool	Tier	Description
retrieve_from_memory(query)	1	Search → short summaries + node IDs
read_detailed_summary(node_id)	2	Lazy-generated detailed summary (cached)
read_raw_memory(node_id)	3	Full original content from vector store

Auto-Ingest

Long tool outputs (> 1500 chars) are automatically ingested into CoMeT. A rolling window ensures agents can immediately see recent results:

• First 2 outputs: Returned raw in full, silently stored in CoMeT
• 3rd output onward: Replaced with node_id reference — agents use read_detailed_summary or read_raw_memory to access

Memory Agent Context

GCRI's Memory Agent receives CoMeT's context window in its prompts, enabling it to leverage in-session knowledge when extracting active constraints and updating external memory on successful task completion.

Configuration (ato)

# comet/config.py
@scope.observe(default=True)
def default(config):
    config.slm_model = 'gpt-4o-mini'
    config.main_model = 'gpt-4o'
    config.compacting.load_threshold = 3
    config.compacting.max_l1_buffer = 5

    # RAG retrieval (enabled when retrieval block exists)
    config.retrieval.embedding_model = 'text-embedding-3-small'
    config.retrieval.vector_backend = 'chroma'
    config.retrieval.vector_db_path = './memory_store/vectors'
    config.retrieval.top_k = 5

@scope.observe()
def local_slm(config):
    config.slm_model = 'ollama/gemma3:4b'

@scope.observe()
def aggressive(config):
    config.compacting.load_threshold = 2
    config.compacting.max_l1_buffer = 3

# Use default
python main.py

# Local SLM + aggressive compacting
python main.py local_slm aggressive

Project Structure

comet/
├── orchestrator.py    # CoMeT main class (3-tier retrieval, document ingestion)
├── sensor.py          # L1 extraction + cognitive load (SLM)
├── compacter.py       # L1→L2 structuring + auto-linking (LLM)
├── storage.py         # JSON key-value store + navigation
├── schemas.py         # MemoryNode, L1Memory, CognitiveLoad, RetrievalResult
├── config.py          # ato scope configuration
├── vector_index.py    # ChromaDB dual-collection vector store (full raw storage)
├── retriever.py       # QueryAnalyzer + ScoreFusion + Retriever
├── consolidator.py    # Dedup + cross-link + tag normalization
└── templates/
    ├── cognitive_load.txt   # Cognitive load judgment prompt
    ├── compacting.txt       # Memory structuring prompt
    ├── l1_extraction.txt    # Fast-layer entity/intent extraction
    └── query_analysis.txt   # Query decomposition prompt