zenmem LLM SDK (Python)

A minimal SDK for building AI features with memory. Start a session, call an LLM, and let it remember things — at whatever scope makes sense: this conversation, this project, or the whole company.

pip install zenmem
pip install "zenmem[all]"       # + openai, anthropic, google-generativeai

from zenmem import ZenmemClient, ZenmemConfig

client = ZenmemClient(ZenmemConfig(
    endpoint="http://203.0.113.10:6636",   # your zenmem deployment — IP:port
    accessToken="<your token>",
    companyCode="WPCORP4812",
))

sessionId = client.startSession()

result = client.callLLM(
    rawPrompt="What should I learn next?",
    provider="OPENAI", model="gpt-4o-mini",
    scope="session", sessionId=sessionId,
    passMemory=True, saveInMemory=True,
)
print(result.output)

client.endSession(sessionId)

That callLLM pulled relevant memory into context, called the model, and wrote a summary of the exchange back — one call.

Requires Python 3.11+. Provider keys come from ZenmemConfig fields or the matching environment variable: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, DEEPSEEK_API_KEY.

endpoint is required — one host:port for everything the SDK talks to. There is no default; every client must say explicitly which deployment it's pointed at.

The basics

Everything in this SDK is one of four things:

Scope

Every memory-touching call takes a scope:

client.addMemory("the user prefers Python", scope="session", sessionId=sessionId)
client.addMemory("our support hours are 9-5 ET", scope="company")

memory = client.fetchMemory("what does the user prefer?", scope="session", sessionId=sessionId)
memory.memoryText   # plain-text block, ready for an LLM prompt
memory.dataNodes    # [DataNode(id, text, score, metadata), ...]

scope="project" reads/writes are partitioned by ZenmemConfig.companyProjectCode, so several projects can share one account without seeing each other's memory. Pass projectId="OTHER_PROJECT" to target a different project for one call.

callLLM

result = client.callLLM(
    rawPrompt="...",                 # or promptId="PROMPT-XXXX"
    provider="OPENAI", model="gpt-4o-mini",   # not needed with promptId
    inputParams={"question": "..."},
    scope="session", sessionId=sessionId,
    passMemory=True,                 # inject memory from `scope` as context
    saveInMemory=True,               # write a summary of this exchange back
)
result.output, result.summary, result.inputTokens, result.outputTokens

Memory retrieval and summarization are non-fatal — if they fail, the model call still goes through without that context.

Sessions

sessionId = client.startSession()          # or startSession("my-own-id")
client.addMemory("...", scope="session", sessionId=sessionId)
client.callLLM(rawPrompt="...", provider="OPENAI", model="gpt-4o-mini",
               scope="session", sessionId=sessionId, passMemory=True)
client.endSession(sessionId)               # promotes its memory to longterm

A closed session id cannot be reused — always start a fresh one for a new conversation.

Module-level API

For scripts that only need one client:

import zenmem

zenmem.init(config)
sessionId = zenmem.startSession()
zenmem.addMemory("...", scope="session", sessionId=sessionId)
zenmem.callLLM(rawPrompt="...", provider="OPENAI", model="gpt-4o-mini")

Errors

All SDK errors derive from ZenmemError, so one except catches everything:

Renamed from VMI

This library was previously vmi-llm-sdk. Nothing breaks on upgrade — the old names are still exported as aliases to the same objects.

Full documentation

The complete reference installs with the package:

import zenmem
print(zenmem.docs_path())

License

MIT. The full text ships in the package as LICENSE.