memharness 0.4.4

Framework-agnostic memory infrastructure for AI agents

memharness

Framework-agnostic memory infrastructure for AI agents

Overview

memharness is a complete memory infrastructure layer for AI agents. It provides:

• 10 Memory Types: Conversational, Knowledge Base, Entity, Workflow, Toolbox, Summary, Tool Log, Skills, File, Persona
• Multiple Backends: PostgreSQL + pgvector, SQLite + sqlite-vss, In-memory
• Framework Agnostic: Works with LangChain, LangGraph, CrewAI, or any custom agent
• Deterministic + AI Operations: Simple ops are deterministic, complex ops use embedded agents
• Self-Exploration Tools: Agents can explore and manage their own memory
• Fully Configurable: All thresholds, schedules, TTLs configurable via YAML or code

Installation

# Core (SQLite backend)
pip install memharness

# With PostgreSQL support
pip install memharness[postgres]

# With embedding support
pip install memharness[embeddings]

# Everything
pip install memharness[all]

Quick Start

from memharness import MemoryHarness

# Initialize with SQLite (development)
memory = MemoryHarness("sqlite:///memory.db")

# Or PostgreSQL (production)
memory = MemoryHarness("postgresql://user:pass@localhost/db")

# Write conversational memory (deterministic)
await memory.add_conversational(
    thread_id="chat_001",
    role="user",
    content="How do I deploy to Kubernetes?"
)

# Write knowledge base (deterministic)
await memory.add_knowledge(
    content="Kubernetes deployment guide...",
    source="k8s-docs",
    metadata={"category": "devops"}
)

# Search knowledge base (semantic search)
results = await memory.search_knowledge(
    query="container orchestration",
    k=5
)

# Get curated context for LLM
context = await memory.assemble_context(
    query="deploy my app",
    thread_id="chat_001",
    max_tokens=4000
)

Memory Types

Type	Storage	Purpose
Conversational	SQL	Chat history per thread
Knowledge Base	Vector	Documents, facts, reference material
Entity	Vector	People, organizations, systems, concepts
Workflow	Vector	Reusable step-by-step patterns
Toolbox	Vector	Tool definitions with VFS discovery
Summary	Vector	Compressed conversations (expandable)
Tool Log	SQL	Tool execution audit trail
Skills	Vector	Learned agent capabilities
File	Hybrid	Document references
Persona	Vector	Agent identity blocks

Embedded Agents

memharness includes specialized agents for complex memory operations:

from memharness import MemoryHarness
from memharness.agents import AgentConfig

memory = MemoryHarness(
    backend="postgresql://...",
    llm=your_llm,  # Optional: for AI-powered operations
    agents=AgentConfig(
        summarizer={"enabled": True, "triggers": [{"condition": "age > 7d"}]},
        entity_extractor={"enabled": True, "on_write": True},
        consolidator={"enabled": True, "schedule": "0 3 * * *"},
        gc={"enabled": True, "schedule": "0 4 * * 0"},
    )
)

Memory Tools (Self-Exploration)

Agents can explore their own memory using built-in tools:

# Get tools for your agent
tools = memory.get_memory_tools()

# Tools include:
# - memory_search: Search across memory types
# - memory_read: Read specific memory by ID
# - memory_write: Write new memory
# - memory_stats: Get memory statistics
# - toolbox_tree: Explore tool VFS
# - toolbox_grep: Search tools by pattern

Configuration

# memharness.yaml
backend: postgresql://localhost/memharness

summarization:
  enabled: true
  triggers:
    - condition: "age > 7d"
      memory_type: conversational
  keep_originals: true
  originals_ttl: 365d

consolidation:
  enabled: true
  schedule: "0 3 * * *"
  similarity_threshold: 0.9

gc:
  enabled: true
  schedule: "0 4 * * 0"
  archive_after: 90d
  delete_after: 365d

memory = MemoryHarness.from_config("memharness.yaml")

Framework Integrations

LangChain

from memharness.integrations.langchain import MemharnessMemory

memory = MemharnessMemory(backend="postgresql://...")
chain = ConversationChain(llm=llm, memory=memory)

LangGraph

from memharness.integrations.langgraph import MemharnessCheckpointer

checkpointer = MemharnessCheckpointer(backend="postgresql://...")
graph = builder.compile(checkpointer=checkpointer)

Documentation

Full documentation: https://ayushsonuu.github.io/memharness

License

MIT License - see LICENSE for details.

Contributing

Contributions welcome! Please read our Contributing Guide.