HipCortex Python SDK (v3.18.2)

Persistent causal topological memory, recursive Bayesian world-model prediction (/worldmodel/rollout), and automatic FSM skill compilation for autonomous AI agents.

Runs against the standalone local Rust binary (http://127.0.0.1:3030) with 0.48–0.61 ms p50 write latency, SHA-256 Merkle audit chains, and Headroom context budgeting (59–88% token savings).

🚀 Installation

pip install hipcortex
hipcortex doctor   # verify server health + version after install
hipcortex install  # wizard: MCP hosts + copies ~/.hipcortex-mcp/server.py (64 tools)
hipcortex channels # honesty matrix (native / mcp / framework / guide)

Version matrix: pip / Cargo / MCP serverInfo / VS Code VSIX = 3.18.2; the TypeScript SDK source is 3.12.0 (npm registry 0.5.2). See docs/channels.md.

What's new in 3.18 — Loop-Ready Contract

An external control loop can now take one snapshot and then follow the changes:

snap = client.cognitive_snapshot(actor="agent")         # bootstrap once
cursor = snap["tx_cursor"]

page = client.snapshot_transitions("agent", since=cursor)
if page["truncated"]:
    snap = client.cognitive_snapshot(actor="agent")     # fell behind the horizon: re-bootstrap
else:
    for t in page["transitions"]:
        ...                                             # upsert by t["intent_id"]
    cursor = page["next_cursor"]

scm = client.snapshot_scm()   # ScmView: nodes, equations, edges, interventions, acyclic, as_of_tx

Settled intents are stamped by the server, ok is true only for a received receipt that reported success, and the feed keeps the newest 4096 settles; horizon_tx is the newest evicted stamp, and a cursor below it gets truncated: true. MCP exposes the same pair as snapshot_transitions and snapshot_scm (64 tools, 7 resources).

⚡ Quick Start: Multi-Tier Memory & World Model Prediction

from hipcortex import HipCortexClient

client = HipCortexClient(base_url="http://127.0.0.1:3030")

# 1. Multi-Tier Memory Ingestion (5 verified memory tiers)
client.add_memory(
    actor="agent",
    action="configured",
    target="jwt_token_ttl=3600",
    record_type="Working",  # Mapped natively to Temporal tier
    priority="high"
)

client.add_memory(
    actor="agent",
    action="established",
    target="All database migrations must run inside transactions",
    record_type="Semantic", # Mapped natively to Symbolic/LongTerm tier
    priority="pinned"
)

# 2. Check SelfModel Execution Capacity Gates before running risky actions
if client.can_execute("rollout"):
    print("Health check passed — engine ready for simulation")

# 3. World Model Trajectory Rollout Prediction (POST /worldmodel/rollout)
rollout_result = client.rollout(
    initial_state={"db_status": "locked", "active_tx": 1},
    actions=["rollback_tx", "release_lock", "retry_migration"]
)
print("Predicted outcome:", rollout_result)

🧠 Proactive CodeAct Harness & Live Beliefs

For autonomous agent harnesses (LangChain, LlamaIndex, CrewAI, AutoGen), query the topological substrate before reasoning:

# Query live causal beliefs and Top-K topological context (Personalized PageRank)
context = client.context(query="database migrations", actor="agent", limit=5)
print("Active Topological Context:\n", context)

# Store multi-hop reasoning or CoT reflection
client.add_memory(
    actor="agent",
    action="reflected",
    target="Migration failed due to missing index on users.email",
    record_type="Causal"
)

🧩 Framework Integrations

LangChain (ConversationBufferMemory drop-in)

from hipcortex.langchain_memory import HipCortexMemory

memory = HipCortexMemory(session_id="session-42", url="http://127.0.0.1:3030")

LlamaIndex (ChatStore)

from hipcortex.llamaindex_storage import HipCortexStorageContext

storage_context = HipCortexStorageContext.from_defaults(client=client)

CrewAI (Remember & Recall Tools)

from hipcortex.adapters.crewai import HipCortexRememberTool, HipCortexRecallTool

tools = [HipCortexRememberTool(client=client), HipCortexRecallTool(client=client)]

AutoGen (Memory Hook)

from hipcortex.adapters.autogen import HipCortexAutoGenMemory

mem = HipCortexAutoGenMemory(client=client, agent_id="researcher")
agent.register_hook("process_message_before_send", mem.on_message_sent)

📊 Server Statistics & GDPR Right-to-Forget

# Get Merkle hash chain status and tier counts
stats = client.stats()
print("Memory statistics:", stats)

# Erase all records for a specific actor (GDPR compliant)
client.forget("agent")