Skip to main content

​
Overview

The Cognitive Memory system enables Kapso to learn from past experiments and provide contextual briefings. It addresses “context stuffing” by intelligently managing what information gets passed to the agent.

​
Architecture

ComponentBackendPurpose
EpisodicStoreWeaviateStores learned insights from past experiments
KnowledgeSearchNeo4j + WeaviateRetrieves domain knowledge and code patterns
Context StateIn-memoryTracks current goal, iteration, and working memory

​
Components

​
EpisodicStore

Stores learned insights from past experiments.
  • Primary: Weaviate vector database for semantic search
  • Fallback: JSON file for persistence without Weaviate
  • Features: Duplicate detection, confidence filtering, automatic pruning

​
CognitiveController

Orchestrates the memory system.
  • Briefing Generation: Creates context packets for agents
  • Insight Extraction: Uses LLM to generalize rules from errors
  • State Management: Persists working memory
  • Decision Making: RETRY, PIVOT, or COMPLETE

​
Usage

​
Basic Usage

from src.memory import CognitiveController, Goal
from src.knowledge.search import KnowledgeSearchFactory

# Initialize with KG search
kg = KnowledgeSearchFactory.create("kg_graph_search")
controller = CognitiveController(knowledge_search=kg)

# Initialize goal
goal = Goal.from_string("Fine-tune LLaMA with LoRA")
controller.initialize_goal(goal)

# Get briefing for agent
briefing = controller.prepare_briefing()
print(briefing.to_string())

# Process experiment result
action, meta = controller.process_result(
    success=False,
    error_message="CUDA OOM: reduce batch size",
    score=0.3,
    feedback="Out of memory error"
)
# Returns: action="retry", meta={"reasoning": "..."}

# Clean up
controller.close()

​
Via Configuration

Enable cognitive memory by setting context manager type: 
context_manager:
  type: "cognitive"
  params:
    max_episodic_insights: 5

​
Data Types

​
Insight

A learned rule with confidence and source tracking: 
@dataclass
class Insight:
    text: str           # The insight content
    confidence: float   # 0.0 to 1.0
    source: str         # Where it came from
    created_at: str     # Timestamp

​
Briefing

Synthesized context packet for the agent: 
@dataclass
class Briefing:
    goal: str
    workflow: str       # From KG (if available)
    heuristics: List[str]
    code_patterns: List[str]
    episodic_insights: List[Insight]

    def to_string(self) -> str:
        # Format for prompt injection

​
Goal

Parsed goal with type classification: 
@dataclass
class Goal:
    text: str
    goal_type: str      # "ml", "data", "web", etc.

    @classmethod
    def from_string(cls, text: str) -> "Goal":
        # Parse and classify

​
Configuration

​
Default Config

defaults:
  episodic:
    embedding_model: "text-embedding-3-small"
    retrieval_top_k: 5
    min_confidence: 0.5
    max_insights: 1000

  controller:
    llm_model: "gpt-4o-mini"
    fallback_models: ["gpt-4.1-mini"]
    max_error_length: 1000

  insight_extraction:
    enabled: true
    max_insight_length: 500
    default_confidence: 0.8

  briefing:
    max_episodic_insights: 5

​
Presets

PresetUse Case
minimalResource-constrained (100 insights)
high_qualityBetter accuracy (large embeddings)
localLocal development (localhost)
dockerDocker deployment
from src.memory.config import CognitiveMemoryConfig

config = CognitiveMemoryConfig.load(preset="high_quality")
controller = CognitiveController(config=config)

​
Insight Extraction

When an experiment fails, insights are extracted: 
# Error: "CUDA OOM when batch_size=32"
# Extracted insight: "For GPU memory issues, reduce batch size below 32"

controller.process_result(
    success=False,
    error_message="CUDA OOM when batch_size=32",
    score=0.0,
)
# Insight stored in EpisodicStore

​
Decision Making

After processing results, the controller decides:
DecisionMeaning
retryTry again with modifications
pivotChange approach significantly
completeTask is done
action, meta = controller.process_result(
    success=True,
    score=0.95,
)
# action = "complete" if score meets threshold

​
Integration

​
With Context Manager

class CognitiveContextManager(ContextManager):
    def __init__(self, ..., knowledge_search):
        self.controller = CognitiveController(knowledge_search=knowledge_search)

    def get_context(self, budget_progress):
        briefing = self.controller.prepare_briefing()
        return ContextData(
            problem=self.problem_handler.get_problem_context(),
            additional_info=briefing.to_string(),
            kg_results=briefing.workflow,
        )

    def should_stop(self):
        return self.controller.last_decision == "complete"

​
Environment Variables

Override settings via environment: 
export COGNITIVE_MEMORY_CONTROLLER_LLM_MODEL=gpt-4-turbo
export COGNITIVE_MEMORY_EPISODIC_EMBEDDING_MODEL=text-embedding-3-large

​
Switching Modes

The cognitive system is opt-in: 
# Legacy (default)
context_manager:
  type: "token_efficient"

# Cognitive (new system)
context_manager:
  type: "cognitive"