LlamaIndex Memory Integration: past dev

past dev is a memory API for AI agents. This package gives a LlamaIndex agent long-term memory in past dev, as a memory block for the LlamaIndex Memory class.

• Get: the block does a past dev Recall. The query is the latest user message. The recalled memory goes into the system message, with the date, the kind and the quoted evidence of each entry.
• Put: when the Memory flushes old messages from its short-term queue, the block does a past dev Ingest of the user and assistant text, as one data point.

Install

pip install llama-index-memory-pastdotdev

Get a key

1. Sign up at past.dev and open a project.
2. Make a project API key. It starts with past_sk_.
3. Set it in the environment as PAST_API_KEY, or give it to the block as api_key.

Use

from llama_index.core.agent.workflow import FunctionAgent
from llama_index.core.memory import Memory
from llama_index.llms.openai import OpenAI
from llama_index.memory.pastdotdev import PastDevMemoryBlock

memory = Memory.from_defaults(
    session_id="session-42",
    token_limit=30000,
    memory_blocks=[
        PastDevMemoryBlock(
            identity=current_user.id,  # your id for the end user, from your own user record
        )
    ],
)

agent = FunctionAgent(tools=[], llm=OpenAI(model="gpt-4.1-mini"))
response = await agent.run("How should you contact me?", memory=memory)

The identity

past dev keeps memory per identity: your id for the end user. Recall needs an identity, and the first ingest for an identity creates it.

Set identity in your code, from your trigger or your user record. Do not let the model choose it. The block has no tool and no prompt that can change the identity, so each user reads and writes only their own memory.

Parameters

What the block sends

• Recall (POST /api/v1/recall): query (the latest user message), identity, and level, maxTokens or limit if you set them.
• Ingest (POST /api/v1/ingest): content (the flushed user and assistant messages as role: text lines; system and tool messages are not sent), identity, timestamp, label, audience if you set it, and metadata with {"source": "llama-index", "sessionId": ...}.

The timestamp is the timestamp of the first flushed message, if your application puts one in message.additional_kwargs["timestamp"]. Otherwise it is the time of the flush.

Credits

Each recall and each ingest spends credits on your past dev plan. The block recalls once each time the agent reads its memory, and ingests once each time the Memory flushes messages.

Errors

An API error raises PastDevError, with the error code, the HTTP status, the message and the X-Request-Id of the request. The block retries a 429 once, after its Retry-After delay. It does not retry other errors, for example 402 out-of-credits or a 400.

Links

• Quickstart
• API reference
• LlamaIndex memory guide
• Tools for the same API: llama-index-tools-pastdotdev

License

MIT. See LICENSE.