saihm-adapters

One memory store your agents own — portable across every model and framework, sealed client-side, and provably erasable.

saihm-adapters gives a Python AI agent a single persistent memory backed by SAIHM: every cell is encrypted inside your process by a bundled Node sidecar — Python never holds a key — so the store is yours, not a vendor's. Point LangChain, LlamaIndex, a RAG retriever, CrewAI, and AutoGen at the same store and the memory travels with your user across frameworks and across model providers. Erase a memory and it is cryptographically shredded (GDPR Art. 17), not merely de-indexed.

• Owned & sovereign — you hold the keys; the server stores ciphertext it cannot read.
• Portable across models & frameworks — one source of truth behind every adapter below.
• Provably erasable — deletion is cryptographic and irreversible, and observable in recall.
• Drop-in — each adapter implements the framework's own memory/history/retriever/storage interface, so it slots into code you already have.

Install

The core client needs only mcp. Each framework adapter is an optional extra, imported lazily — install only what you use:

pip install saihm-adapters                 # core client (any Python app)
pip install "saihm-adapters[langchain]"    # + LangChain chat history
pip install "saihm-adapters[langgraph]"    # + LangGraph BaseStore
pip install "saihm-adapters[llamaindex]"   # + LlamaIndex memory
pip install "saihm-adapters[rag]"          # + LlamaIndex retriever for RAG
pip install "saihm-adapters[crewai]"       # + CrewAI storage backend
pip install "saihm-adapters[autogen]"      # + AutoGen memory
pip install "saihm-adapters[all]"          # every adapter

Node.js ≥ 20 must be on PATH. The wheel ships only the tiny sidecar sources (server.mjs / sandbox.mjs + a pinned lockfile); its Node dependencies are installed once into a per-user cache on first use (npm ci). Python performs no cryptography — the Node sidecar seals every cell. For air-gapped or CI installs, pre-provision the sidecar and set SAIHM_SIDECAR_DIR; SAIHM_SKIP_BOOTSTRAP=1 forbids any network install.

Quickstart

The core client works on its own — remember, recall, and erase, from any Python app:

from saihm_adapters import SaihmMemoryClient

client = SaihmMemoryClient()          # reads SAIHM_* env for your account
cid = client.remember("Ada prefers metric units.")
for cell in client.recall("units"):
    print(cell.content)
client.forget(cid)                    # cryptographic erase

Per-framework adapters

Each adapter accepts a shared client= so a whole app (or a whole crew) reads and writes one store:

# LangChain — a BaseChatMessageHistory
from saihm_adapters import SaihmChatMessageHistory
history = SaihmChatMessageHistory(client=client)

# LangGraph — a BaseStore for long-term, cross-thread memory
from saihm_adapters import SaihmStore
store = SaihmStore(client=client)

# RAG — a LlamaIndex BaseRetriever over a corpus you own
from saihm_adapters import SaihmRetriever
retriever = SaihmRetriever(client=client, similarity_top_k=3)

# CrewAI — a Storage backend for ExternalMemory
from saihm_adapters import SaihmStorageBackend
backend = SaihmStorageBackend(client=client)

# AutoGen — an autogen_core.memory.Memory
from saihm_adapters.autogen_memory import SaihmMemory as AutoGenMemory
memory = AutoGenMemory(client=client)

# LlamaIndex — a BaseMemory
from saihm_adapters.llamaindex_memory import SaihmMemory as LlamaIndexMemory
memory = LlamaIndexMemory.from_defaults(client=client)

Why SaihmMemory is imported from a submodule: AutoGen and LlamaIndex each define a class named SaihmMemory, and they are different classes. Importing from saihm_adapters.autogen_memory / saihm_adapters.llamaindex_memory keeps the two unambiguous. The other adapter classes have unique names and import straight from the top level.

How it works

Python drives a bundled Node MCP sidecar over stdio. Your plaintext is sealed (ML-KEM-768 + AES-256-GCM, signed with ML-DSA-65) inside that sidecar before anything leaves the process; the SAIHM service stores only ciphertext. Recall decrypts in-process. Because the store is keyed to you, the same memory is reachable from any framework and any model — and an erase removes the key material, so the data cannot be recovered.

Live demos

Runnable notebooks and end-to-end examples: https://citw2.github.io/saihm-demos/

License

Apache-2.0.