mubit-crewai 0.6.0

CrewAI memory backend backed by MuBit memory engine

mubit-crewai

CrewAI memory backend backed by MuBit.

This adapter provides a class that satisfies CrewAI's crewai.memory.storage.backend.StorageBackend protocol (verified against crewai>=1.14.6,<2), routing all memory operations through MuBit's control plane. Also includes an extended MubitCrewMemory class with MAS coordination helpers and the outcome-attribution loop.

Install

pip install mubit-crewai[crewai]

Basic usage

Drop-in StorageBackend

from mubit_crewai import MubitStorage
from crewai import Crew
from crewai.memory import Memory

storage = MubitStorage(api_key="mbt_...", session_id="crew-run-1")
crew = Crew(
    agents=[...],
    memory=Memory(storage=storage),
)

Extended MuBit features

from mubit_crewai import MubitCrewMemory, extract_entry_ids

memory = MubitCrewMemory(api_key="mbt_...", session_id="crew-run-1")
crew = Crew(agents=[...], memory=memory.as_crew_memory())

# Extended:
context = memory.get_context("What do we know about the customer?")
memory.checkpoint("Finished research phase")
memory.handoff("researcher", "writer", "Here are the findings...")

Outcome-attribution loop (v0.7.0)

Close the recall -> act -> credit loop so the entries that helped get reinforced, and process rewards land per step:

result = memory.recall("How did we resolve the last billing dispute?")
entry_ids = extract_entry_ids(result)       # reference_ids of recalled evidence
# ... the crew acts on the recalled context ...
memory.record_outcome(
    "task-123", "success",
    entry_ids=entry_ids,                     # credit the entries that helped
    verified_in_production=True,             # boost lessons confirmed in live use
)
memory.record_step_outcome("plan", outcome="success")  # per-step process reward

StorageBackend interface

Implements the full StorageBackend protocol — save, search, delete, update, get_record, list_records, get_scope_info, list_scopes, list_categories, count, reset, plus async asave/asearch/adelete.

Method	MuBit mapping
save(records)	control.ingest (intent from categories, idempotency keyed by record id)
search(query_embedding, ...)	control.query, returns (MemoryRecord, score) tuples
delete(...) / reset()	local-index delete / control.delete_run

A legacy save(value, metadata, agent_id) / search(query, limit, score_threshold) shim is retained for older callers.

Intent inference

The memory_type metadata key maps to MuBit intents:

memory_type	MuBit intent
short_term	trace
long_term	lesson
entity	fact

MuBit extension methods

• remember() / recall() — direct write/read (idempotent writes, citations on recall)
• get_context() — assembled context retrieval
• checkpoint() — save memory state snapshot
• record_outcome() — outcome feedback for tasks (with entry_ids + verified_in_production)
• record_step_outcome() — per-step process reward
• learned() — persist a confirmed lesson
• surface_strategies() — strategy clusters from lessons
• register_agent() — MAS agent registration
• handoff() — agent-to-agent handoff
• diagnose() — failure-path lesson surfacing

Module helpers extract_entry_ids(result) and extract_citations(result) are re-exported for closing the attribution loop on a recall() result.

Config

Parameter	Default	Purpose
endpoint	http://127.0.0.1:3000	MuBit HTTP endpoint
api_key	""	MuBit API key
session_id	"default"	Session/run scope
agent_id	"crewai"	Default agent ID

For tests, inject mubit_client directly.

Development

PYTHONPATH=sdk/python/mubit-sdk/src:integrations/python \
python3 -m unittest integrations.python.mubit_crewai.tests.test_storage -v

License

Apache-2.0