condensate 0.2.4

Official Python Client for Condensate Memory System

Condensate Python SDK

Official Python client for Condensate — the open-source Agent Memory System.

Installation

pip install condensate

Requires Python 3.9+.

Quick Start

from condensate import CondensateClient

client = CondensateClient(
    base_url="http://localhost:8000",
    api_key="sk-your-api-key"
)

# Store a memory
client.store_memory(
    content="The team decided to use PostgreSQL for the primary store.",
    type="episodic",
    metadata={"source": "meeting", "project": "infra-v2"}
)

# Retrieve relevant memories
result = client.retrieve("What database did we choose?")
print(result["answer"])
print(result["sources"])   # list of episodic item IDs
print(result["strategy"])  # "recall" | "research" | "meta"

Configuration

Environment Variables

Set these before initialising the client, or pass them directly:

Variable	Description	Default
CONDENSATE_URL	Base URL of your Condensate server	http://localhost:8000
CONDENSATE_API_KEY	API key (create one in the admin dashboard)	—

import os
from condensate import CondensateClient

client = CondensateClient(
    base_url=os.environ["CONDENSATE_URL"],
    api_key=os.environ["CONDENSATE_API_KEY"]
)

Connecting to a Remote Server

client = CondensateClient(
    base_url="https://memory.yourcompany.com",
    api_key="sk-prod-xxxx"
)

API Reference

store_memory(content, type, metadata)

Stores a raw episodic item and runs the full condensation pipeline (NER → entity extraction → assertion creation).

client.store_memory(
    content="Alice approved the Q3 roadmap.",
    type="episodic",          # episodic | note | event
    metadata={
        "source": "slack",
        "channel": "#product"
    }
)

retrieve(query)

Routes the query through the Memory Router (vector search + optional graph traversal) and returns a synthesised answer.

result = client.retrieve("What did Alice approve?")
# {
#   "answer": "Alice approved the Q3 roadmap.",
#   "sources": ["<uuid>", ...],
#   "strategy": "recall"
# }

add_item(item) (low-level)

Directly posts an EpisodicItemCreate payload to the v1 API.

from condensate import EpisodicItem
import uuid

client.add_item(EpisodicItem(
    project_id=str(uuid.uuid4()),
    source="api",
    text="User prefers dark mode."
))

Orchestration Lifecycle Hooks (Symphony Orchestration)

To support stateful orchestration in Symphony-like multi-agent networks, the SDK exposes the CondensateOrchestrationHooks helper. It maps lifecycle transitions to standard episodic memories, utilizing the /api/v1/episodic endpoint and a robust metadata-driven approach:

from condensate import CondensateClient, CondensateOrchestrationHooks

client = CondensateClient(base_url="http://localhost:8000", api_key="sk-your-key")
hooks = CondensateOrchestrationHooks(client)

# When an agent execution session starts
hooks.on_agent_started(task_id="linear-ticket-101", agent_id="agent-developer-1", agent_role="developer")

# Checkpoint/Suspend state for fault-tolerance or hand-offs
state_dump = {
    "current_file": "src/server/v1_api.py",
    "cursor_position": 142,
    "progress_percent": 65.5
}
hooks.on_agent_suspended(task_id="linear-ticket-101", agent_id="agent-developer-1", state_dump=state_dump)

# Resume execution
hooks.on_agent_resumed(task_id="linear-ticket-101", agent_id="agent-developer-1")

# If an agent encounters a fatal error / crashes
hooks.on_agent_crashed(task_id="linear-ticket-101", agent_id="agent-developer-1", error="Database connection timeout", state_dump=state_dump)

# On successful task completion
hooks.on_agent_completed(task_id="linear-ticket-101", agent_id="agent-developer-1", final_findings="All security patches applied successfully.")

Under the Hood: Metadata-Driven Design

Every event emitted by CondensateOrchestrationHooks is sent to the /api/v1/episodic endpoint with source set to "orchestrator" and the payload's metadata dictionary populated with event_type, task_id, and agent_id. Because Condensate automatically scopes all queries and retrievals to the API key's project, lifecycle states are 100% tenant-isolated, allowing clean, trace-based analysis of multi-agent execution lifecycles.

CLI

The package ships a condensate CLI:

# Store a memory from stdin
echo "Deploy to prod on Friday" | condensate store --type episodic

# Retrieve
condensate retrieve "When are we deploying?"

Set CONDENSATE_URL and CONDENSATE_API_KEY in your shell before using the CLI.

Self-Hosting

See the main README for Docker Compose setup. The server must be running before the SDK can connect.

Development

cd sdks/python
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest