crewai-pastdotdev

past dev is a memory API for AI agents. This package connects it to CrewAI: your crews keep what they learn in past dev and recall it in later runs, with the date and the source of every memory.

It gives you two ways to use past dev:

• PastDevMemory: crew memory. Pass it as memory= to a Crew or an Agent. CrewAI recalls before each task and saves each task result. A save is a past dev Ingest and a search is a past dev Recall.
• PastDevRecallTool and PastDevIngestTool: tools that an agent calls when it decides to. The model gives only the query or the content.

Website · Quickstart · API reference

Install

pip install crewai-pastdotdev

The package needs Python 3.10 to 3.13 and CrewAI 1.14.6 or later.

Set the API key

Make a project API key in the past dev console. It starts with past_sk_.

export PAST_API_KEY="past_sk_..."

If you run past dev in your own cloud, also set the API host. The default is https://api.past.dev.

export PAST_API_HOST="https://past.example.com"

You can also give api_key= and api_host= to each class.

The identity

Every memory in past dev belongs to an identity: the id of the person or account the memory is about, for example the user id in your application. Recall requires an identity. Ingest makes the identity the first time it sees it.

Your code sets the identity, never the model. Take it from the user record or the trigger of the run, and give it to the constructor. The tools have no identity argument, so an agent cannot read or write the memory of another person. Make one memory or one tool for each user.

Use past dev as crew memory

from crewai import Agent, Crew, Task
from crewai_pastdotdev import PastDevMemory

memory = PastDevMemory(identity=current_user.id)

writer = Agent(
    role="Account manager",
    goal="Write the weekly report the way the customer wants it",
    backstory="You know the customer well.",
)
report = Task(
    description="Write this week's report for the customer.",
    expected_output="A report in the format the customer prefers.",
    agent=writer,
)

crew = Crew(agents=[writer], tasks=[report], memory=memory)
crew.kickoff()

What happens in a run:

1. Before each task, CrewAI recalls with the task description. The recalled memories go into the agent's prompt.
2. After each task, CrewAI saves the task and its result. PastDevMemory ingests them as one data point. past dev extracts the memories on its own side, so no extra LLM call runs in your crew.
3. CrewAI also gives the agents its own "Search memory" and "Save to memory" tools. They go through past dev too.

You can also call the memory directly:

record = memory.remember("The customer wants the report on Mondays.")
for match in memory.recall("When does the customer want the report?"):
    print(match.record.content, match.record.metadata["occurred_at"])
memory.forget(record_ids=[record.id])  # Delete data point

Settings

Argument	Default	What it does
identity	required	The identity for every Ingest and Recall.
api_key	PAST_API_KEY	The project API key.
api_host	PAST_API_HOST, else https://api.past.dev	The API host.
audience	none	An audience slug for ingested memories. Without it, a memory is visible to the whole project.
label	none	A short title stored with each data point.
recall_level	low	The evidence budget of a recall: low, medium, high or extra-high.
recall_max_tokens	none	An exact evidence budget in tokens, instead of recall_level.
read_only	False	Recall only. Nothing is ingested.

How CrewAI concepts map to past dev

CrewAI	past dev
remember(), remember_many()	Ingest (POST /api/v1/ingest). The source id is stable for the same identity and content, so a repeated save is free.
recall()	Recall (POST /api/v1/recall). limit is the number of memories, at most 200.
forget(record_ids=[...])	Delete data point (DELETE /api/v1/data-points/{sourceId}).
MemoryMatch.score	1 / rank. past dev ranks the results; it does not return a score.
Scopes, categories, agent role	Kept in the data point's metadata (crewai_scope, crewai_categories, crewai_agent_role). Recall does not filter on them.
tree(), info(), list_records(), update(), reset()	Not available: past dev keeps memory on its servers. Use the past dev console to browse. To change a fact, remember the new one: past dev marks the earlier statement as superseded.

Use past dev as agent tools

from crewai import Agent
from crewai_pastdotdev import PastDevIngestTool, PastDevRecallTool

agent = Agent(
    role="Support agent",
    goal="Help the customer with what you know about them",
    backstory="You remember every earlier conversation.",
    tools=[
        PastDevRecallTool(identity=current_user.id),
        PastDevIngestTool(identity=current_user.id),
    ],
)

Tool	The model gives	Other settings (in the constructor)
PastDevRecallTool ("past dev Recall")	query	identity, limit (10), level, api_key, api_host
PastDevIngestTool ("past dev Ingest")	content	identity, audience, label, metadata, api_key, api_host

The recall tool returns the memories as text: rank, date, content, and the quotations from their sources. If a memory was superseded by a later statement, the text says so.

Credits and errors

Every Ingest and every Recall spends credits on your past dev plan. The connection check (PastDevClient().check_connection(), which reads the project's audiences) and Delete data point are free.

An error from the API raises PastDevError with the API's code (for example out-of-credits), the HTTP status, the message and the request_id. Give the request id to past dev support when you ask about a call. The tools return the error to the agent as text, and tell it not to retry when a retry cannot help.

The client sends a request again only when the API answers 429 for the rate limit, after the wait the API asks for (two times at most). It never retries 402 out-of-credits, a monthly cap, or a 400.

The HTTP client

PastDevClient is a small client for the calls this package uses. You can use it on its own:

from crewai_pastdotdev import PastDevClient

with PastDevClient() as past:
    past.check_connection()
    past.ingest("The customer moved to Lyon.", identity="user-42", id="crm-note-17")
    page = past.recall("Where does the customer live?", identity="user-42", limit=5)
    past.delete_data_point("crm-note-17")
    audiences = past.list_audiences()
    identities = list(past.list_identities())

Development

python -m venv .venv && . .venv/bin/activate
pip install -e ".[test]"
pytest

The tests use httpx.MockTransport and a scripted LLM, so they need no key and no network. One live test runs a real Ingest, Recall and Delete data point when PAST_API_KEY is set. It spends a few credits.

License

MIT. See LICENSE.