Memory¶

OpenDataSci keeps conversation history as rolling turn summaries rather than raw messages. Completed turns are summarised and discarded; only the current in-progress turn is kept in full.

Concepts¶

Class	What it represents
`ChatTurnSummary`	A compressed record of a single completed turn.
`ChatHistoryCompaction`	A further-folded narrative produced by `Agent.compact_chat_history()`.
`ChatTurnContext`	The assembled messages fed to the LLM at the start of each turn — summaries, compaction recall, current plan, and the ongoing turn.

Inspecting memory from graph state¶

Turn summaries and the compaction record live in AgentState and can be read from a graph snapshot:

snapshot = agent.graph.get_state({"configurable": {"thread_id": session_id}})
turn_summaries = snapshot.values.get("turn_summaries", [])
for record in turn_summaries:
    print(record.to_content())

compaction = snapshot.values.get("chat_history_compaction")
if compaction is not None:
    print(compaction.to_content())

Message types¶

opendatasci.memory.messages defines typed LangChain message subclasses. Each carries a created_at timestamp. The principal user-facing type is UserMessage, which marks a message as originating from the user and flags interrupt replies via is_input_on_interrupt.

Class	Base	Purpose
`UserMessage`	`HumanMessage`	A message from the user.
`CompactionMessage`	`HumanMessage`	A harness message carrying a compacted history recall block.
`AgentToAgentMessage`	`HumanMessage`	A message sent from the orchestrator to a sub-agent.
`SummaryMessage`	`HumanMessage`	A harness message carrying a turn-summary recall block.
`PlanMessage`	`HumanMessage`	A harness message carrying the current session plan.
`AgentMessage`	`AIMessage`	A message produced by the LLM agent.

Reference¶

UserMessage ¶

Bases: HumanMessage, RenderableMessageMixin['UserMessage']

A message that originated directly from the user.

CompactionMessage ¶

Bases: HumanMessage, RenderableMessageMixin['CompactionMessage']

A harness message carrying a compacted history recall block.

AgentToAgentMessage ¶

Bases: HumanMessage, RenderableMessageMixin['AgentToAgentMessage']

A message sent from one agent to another (e.g. orchestrator to ConcurrentWorkerAgent).

SummaryMessage ¶

Bases: HumanMessage, RenderableMessageMixin['SummaryMessage']

Harness-constructed message carrying a turn-summary recall.

PlanMessage ¶

Bases: HumanMessage, RenderableMessageMixin['PlanMessage']

Harness-constructed message carrying the current session plan.

AgentMessage ¶

Bases: AIMessage

A message produced by the LLM agent.

ChatTurnContext `dataclass` ¶

ChatTurnContext(messages: list[BaseMessage], turn_summaries: list[ChatTurnSummary], chat_history_compaction: ChatHistoryCompaction | None)

The assembled messages for a single LLM call.

Attributes:

Name	Type	Description
`messages`	`list[BaseMessage]`	Compaction recall (if any), turn-summary recall messages, the current plan (if any), and the ongoing turn's messages — all rendered for the LLM.
`turn_summaries`	`list[ChatTurnSummary]`	Updated rolling summary list to write back to agent state.
`chat_history_compaction`	`ChatHistoryCompaction \| None`	Updated compaction to write back to agent state. `None` means either no compaction exists or it was cleared because the summary window became full and the compaction is no longer needed.

ChatTurnSummary `dataclass` ¶

ChatTurnSummary(turn_start_timestamp: datetime, turn_end_timestamp: datetime, user_message_summary: str, actions_summary: str, agent_response_summary: str)

Bases: LLMDigestibleMixin

Summary of a single completed conversation turn.

ChatHistoryCompaction `dataclass` ¶

ChatHistoryCompaction(compacted_at: datetime, timespan: tuple[datetime, datetime] | None, content: str)