File size: 3,382 Bytes
5e028bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | <h1 align="center"> StructMem: Structured Memory for Long-Horizon Behavior in LLMs </h1>
---
**StructMem**, a structure-enriched hierarchical memory framework that preserves event-level bindings and induces cross-event connections. By temporally anchoring dual perspectives and performing periodic semantic consolidation, StructMem improves temporal reasoning and multi-hop performance.
<div align=center><img src="./figs/StructMem.png" width="100%" height="60%" /></div>
## Key Contributions
### 1. Event-Level Extraction Mode
StructMem introduces an **event extraction mode** that goes beyond flat factual entries by capturing:
- **Factual components**: Core information units (who, what, when, where)
- **Relational components**: Interpersonal dynamics, causal influences, and temporal dependencies.
- **Temporal binding**: Event sequences and causal relationships
This mode is particularly effective for narrative-heavy conversations and time-sensitive scenarios.
**Usage:**
```python
config = {
# ... other configs ...
"extraction_mode": "event", # 'flat' or 'event'
# ... other configs ...
}
```
### 2. Cross-Event Summarization
StructMem implements a **hierarchical summarization mechanism** that:
- **Semantic Event Connections**: Buffers related events within time windows
- **Memory Consolidation through Synthesis**: Generates cross-event summaries stored separately from detailed memories
**Usage:**
```python
# Generate summaries after building memories
summary_result = lightmem.summarize(
retrieval_scope="global",
time_window=3600, # seconds
top_k=15, # number of seed memories
process_all=True
)
```
**Configuration:**
```python
config = {
# ... other configs ...
"summary_retriever": {
"model_name": "qdrant",
"configs": {
"collection_name": "my_summaries",
"embedding_model_dims": 384,
"path": "./my_summaries",
}
},
# ... other configs ...
}
```
## Usage Example
See the complete example in [`experiments/locomo/`](./experiments/locomo/) for building and evaluating StructMem on the LoCoMo dataset.
### Quick Start
```python
from lightmem.memory.lightmem import LightMemory
# Configure StructMem
config = {
# ... base LightMem configs ...
# StructMem-specific settings
"extraction_mode": "event",
"summary_retriever": {
"model_name": "qdrant",
"configs": {
"collection_name": "my_summaries",
"embedding_model_dims": 384,
"path": "./my_summaries",
}
},
# ... other configs ...
}
# Initialize
lightmem = LightMemory.from_config(config)
# Add memories with event extraction, providing both factual and relational prompts.
lightmem.add_memory(
messages=turn_messages,
METADATA_GENERATE_PROMPT={
"factual": LoCoMo_Event_Binding_factual,
"relational": LoCoMo_Event_Binding_relational
},
force_segment=True,
force_extract=True
)
# Generate cross-event summaries
lightmem.summarize(
retrieval_scope="global",
time_window=3600,
top_k=15,
process_all=True
)
```
## Experimental Results
StructMem demonstrates improved performance on long-context conversation benchmarks. For complete results and configurations, see [LoCoMo Results](./experiments/locomo/readme.md#structmem-results)
|