khwan (Python client)

The Khwan hosted client — a thin HTTP wrapper with no engine code. Khwan is a memory layer (memory + constitutional identity + coherence + learning) that runs on our server; you bring your own model.

Khwan never generates text. It is a pure AI-memory layer — you always call your own model. The only loop is prepare → your model → record.

pip install khwan

The memory loop — you call your own model

from khwan import Khwan

kw = Khwan(api_key="kwk_live_xxx", user_id="alice")

turn   = kw.prepare("remember I prefer short answers in Thai")  # Khwan builds context, no LLM
answer = your_model(turn.messages)                              # YOUR model + key
kw.record(turn, answer)                                          # Khwan persists + learns

# `record` waits by default, on purpose: `prepare` for the next turn reads what
# has been written, so a record still in flight drops this turn from the next
# turn's context — only under load, which makes it read as flaky memory rather
# than as a race. Skip the wait when the turn is the last one:
kw.record(turn, answer, background=True)                         # → {"queued": True}

# The send runs on a daemon thread, and the interpreter does not wait for those.
# In a CLI, a serverless handler, or any script that ends soon after its last
# turn, that write can be killed mid-flight — no error, the turn simply never
# learned. Wait for it before you exit:
kw.flush()                                                       # → how many were in flight

A flush() also runs automatically at interpreter exit, bounded to five seconds, so forgetting the call costs latency rather than the turn.

On an event loop

Every agent framework worth integrating is async, and a blocking client on an event loop either stalls it or grows a thread pool to hide the stall. AsyncKhwan is the same loop, the same retry rules — they live at module level, so the two clients cannot drift — and the same errors.

pip install "khwan[async]"

from khwan import AsyncKhwan

# Holds one connection pool, so keep it open rather than building one per turn.
async with AsyncKhwan(api_key="kwk_live_xxx", core="acme", user_id="Web") as kw:
    turn   = await kw.prepare("what did we decide about billing?")
    answer = await your_model(turn.messages)
    await kw.record(turn, answer)

    await kw.record(turn, answer, background=True)   # → {"queued": True}

background=True schedules the write and returns immediately; aclose() — which async with calls for you — waits for anything still in flight, so a fire-and- forget record is not lost when the process ends.

What the brain already knew

prepare returns the raw turns it retrieved and the rules synthesis has distilled from many past turns. Both are already inside turn.messages; they are also exposed so a caller building its own context — a recall tool, a subagent brief — can take the distilled rules without replaying the whole prompt.

turn.lessons   # ["Answer in Thai.", …]  standing rules
turn.sources   # the raw turns retrieved for THIS turn, each with a similarity

Retrieval applies a relevance floor, so an empty sources is an answer: the brain has nothing close to this question. Read it as "not known here" rather than reaching for whichever memory was nearest.

Gate the answer, review what it learned

v = kw.verify(turn, draft)          # score a draft BEFORE you ship it
if not v["ok"]:
    ...                             # regenerate, or route to a human

for l in kw.lessons():              # the standing rules it distilled
    print(l["response_text"], "←", l["source_link"])
kw.delete_lesson(bad_id)            # the only negative signal in the system

Own the learning step

prepare → your model → record covers answering. The same shape covers learning — Khwan clusters the turns, your model writes the rule, so no packet text reaches a provider Khwan chose:

kw.synthesize(distill=lambda system, prompt: my_llm(system, prompt))

turn.messages is a standard [{role, content}] array with Khwan's value baked into the system prompt (learned lessons + constitution + retrieved memory + coherence). your_model is just your normal LLM call:

import anthropic
client = anthropic.Anthropic(api_key="sk-ant-...")  # your key, Khwan never sees it

def your_model(messages):
    system = next((m["content"] for m in messages if m["role"] == "system"), "")
    chat   = [m for m in messages if m["role"] != "system"]
    r = client.messages.create(model="claude-sonnet-4-6", max_tokens=1024,
                               system=system, messages=chat)
    return r.content[0].text

Isolated cores

One account can hold many cores — fully separate brains, each with its own memory, identity, and learning. Point a client at one with core:

test    = Khwan(api_key="kwk_live_xxx", user_id="alice", core="test")
client1 = Khwan(api_key="kwk_live_xxx", user_id="alice", core="client1")

kw.cores()   # list the account's cores (the default core is included)

test and client1 never share memory. Omit core for the account's default brain.

On-prem

Same code, point at your instance:

kw = Khwan(api_key="kwk_...", user_id="alice",
               base_url="https://khwan.internal.acme.com")

memory=/embedder= are server-managed and rejected here — they exist only in the on-prem engine, shipped under license.