memharness

Framework-agnostic memory infrastructure for AI agents

memharness provides the memory layer for AI agents — persistent, searchable, typed memory with lifecycle management. Works with any agent framework.

📖 Documentation · 📦 PyPI · 🐛 Issues

Install

pip install memharness

# With PostgreSQL
pip install memharness[postgres]

# With HuggingFace embeddings
pip install memharness[embeddings]

Quick Start

from memharness import MemoryHarness

async with MemoryHarness("sqlite:///memory.db") as harness:
    # Store memories
    await harness.add_conversational("thread-1", "user", "I work at SAP")
    await harness.add_knowledge("Python 3.13 has free-threading", source="docs")
    await harness.add_entity("Alice", "PERSON", "Engineer at Acme Corp")

    # Search semantically
    results = await harness.search_knowledge("concurrent programming")

    # Assemble context for any LLM
    from memharness.agents import ContextAssemblyAgent
    ctx = ContextAssemblyAgent(harness)
    context = await ctx.assemble("Tell me about Python", thread_id="thread-1")
    messages = context.to_messages()  # list[BaseMessage] for LangChain

Memory Types

8 types covering the full agent memory taxonomy:

Read-Only Memory Tools

Give any agent read-only access to its memory:

from memharness.tools import get_read_tools

tools = get_read_tools(harness)  # Returns 5 LangChain BaseTool instances

Embedded Agents

4 meta-agents that manage memory autonomously:

Summarization

After summarization, context loads summary + recent messages only — not all messages:

Before: [msg1, msg2, ... msg50]           ← 50 messages in context
After:  [Summary of msg1-40] + [msg41-50] ← 1 summary + 10 recent

Configurable via max_tokens (default 4000) and summarize_threshold (default 80%).

Backends

Docker

# PostgreSQL + pgvector
docker compose up -d

Use with LangChain

from langchain.agents import create_agent
from memharness.tools import get_read_tools

agent = create_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=get_read_tools(harness),
)

See the full LangChain usage guide with middleware examples.

Configuration

harness = MemoryHarness(
    backend="sqlite:///memory.db",
    embedding_fn=my_embedding_function,  # default: hash-based
)

# Or with HuggingFace embeddings
from memharness.core.embedding import create_huggingface_embedding_fn
harness = MemoryHarness(
    backend="sqlite:///memory.db",
    embedding_fn=create_huggingface_embedding_fn("all-MiniLM-L6-v2"),
)

Documentation

Full docs at ayushsonuu.github.io/memharness

Contributing

Contributions welcome! See CONTRIBUTING.md.

License

MIT — see LICENSE.