sdk-python

TinyHuman Neocortex SDK

A persistent memory layer for AI applications. Neocortex lets your AI agents store, retrieve, and use context across conversations -- so they remember what matters.

Built on the TinyHumans API.

Install

pip install tinyhumansai

Requires Python 3.9+. The only runtime dependency is httpx.

Running locally

From the package directory (packages/sdk-python), install the project and optional dependency groups with uv:

uv sync --group dev --group examples

• --group dev — test and lint tools (pytest, pytest-asyncio, mypy, ruff).
• --group examples — python-dotenv for running the example script.

Then run the example:

# Copy .env.example to .env and set TINYHUMANS_TOKEN, TINYHUMANS_MODEL_ID, OPENAI_API_KEY (for recall_with_llm).
uv run python example.py

# Optional: print outbound SDK requests for debugging
TINYHUMANSAI_LOG_LEVEL=DEBUG uv run python example.py

To install only the package and examples (no dev tools): uv sync --group examples.

Quick start

import tinyhumansai as api

client = api.TinyHumanMemoryClient("YOUR_APIKEY_HERE")

# Store a single memory
client.ingest_memory(
    item={
        "key": "user-preference-theme",
        "content": "User prefers dark mode",
        "namespace": "preferences",
        "metadata": {"source": "onboarding"},
    }
)

# Fetch relevant memory context, then ask a LLM something from it
ctx = client.recall_memory(
    namespace="preferences",
    prompt="What is the user's preference for theme?",
)

response = client.recall_with_llm(
    prompt="What is the user's preference for theme?",
    api_key="OPENAI_API_KEY",
    context=ctx.context,
)
print(response.text) # The user prefers dark mode

Core concepts

Memory items are the basic unit of storage. Each item has:

Field	Required	Description
key	yes	Unique identifier within a namespace (used for upsert/dedup)
content	yes	The memory text
namespace	yes	Scope for organizing items
metadata	no	Arbitrary dict for tagging/filtering
created_at	no	Unix timestamp in seconds
updated_at	no	Unix timestamp in seconds

Namespaces let you organize memories by category (e.g. "preferences", "conversation-history", "user-facts").

Context is a pre-formatted string built from your stored memories, ready to inject into any LLM prompt as system context.

API reference

TinyHumanMemoryClient

client = api.TinyHumanMemoryClient(
    token="your-api-key",       # Required. TinyHumans API key.
    model_id="neocortex-mk1",   # Required. Model identifier.
    base_url="https://...",     # Optional. Override API base URL.
)

The client supports the context-manager protocol for automatic cleanup:

with api.TinyHumanMemoryClient(token="...", model_id="...") as client:
    ctx = client.recall_memory(namespace="preferences", prompt="User preferences", num_chunks=10)

ingest_memory

Upsert a single memory item. The item is deduped by (namespace, key) -- if a match exists, it is updated; otherwise a new item is created.

result = client.ingest_memory(
    item={
        "key": "fav-color",
        "content": "User's favorite color is blue",
        "namespace": "preferences",
    }
)
print(result.ingested, result.updated, result.errors)

With the MemoryItem dataclass:

from tinyhumansai import MemoryItem

result = client.ingest_memory(
    item=MemoryItem(key="fav-color", content="Blue", namespace="preferences")
)

ingest_memories

Upsert multiple memory items in one call. Items are deduped by (namespace, key).

result = client.ingest_memories(
    items=[
        {"key": "fav-color", "content": "Blue", "namespace": "preferences"},
        {"key": "fav-food", "content": "Pizza", "namespace": "preferences"},
    ]
)
print(result.ingested, result.updated, result.errors)

recall_memory

Fetch relevant memory chunks using a prompt and return them as an LLM-friendly context string. The API uses the prompt to retrieve the most relevant chunks from the namespace.

# Fetch up to 10 chunks relevant to the prompt
ctx = client.recall_memory(
    namespace="preferences",
    prompt="What is the user's favorite color?",
    num_chunks=10,
)
print(ctx.context)  # Formatted string
print(ctx.items)    # List of ReadMemoryItem objects
print(ctx.count)    # Number of items

# Optional: fetch more or fewer chunks
ctx = client.recall_memory(namespace="preferences", prompt="User preferences", num_chunks=5)

# Optional: filter by specific key(s) instead of prompt-based retrieval
ctx = client.recall_memory(namespace="preferences", prompt="", key="fav-color", num_chunks=10)

delete_memory

Delete all memories in a namespace. The current API does not expose key-scoped deletes.

# Delete all memories in a namespace
client.delete_memory(namespace="preferences", delete_all=True)

recall_with_llm (optional)

Query an LLM provider with your stored context injected -- no extra SDK dependencies needed. Supports OpenAI, Anthropic, and Google Gemini out of the box, plus any OpenAI-compatible endpoint.

ctx = client.recall_memory(namespace="preferences", prompt="User preferences", num_chunks=10)

# OpenAI
response = client.recall_with_llm(
    prompt="What is the user's favorite color?",
    provider="openai",
    model="gpt-4o-mini",
    api_key="your-openai-key",
    context=ctx.context,
)
print(response.text)