dragon.ai.agent.config.memory_config.MemoryConfig

class MemoryConfig

Bases: object

Configures how an agent manages its conversation history (work memory).

The ToolDispatcher accumulates messages (system prompt, user task, assistant tool-call requests, tool results) in a growing list and sends the entire list to the LLM on every iteration. Without memory management, this list grows without bound and can exceed the model’s context window.

MemoryConfig provides three strategies:

• MemoryStrategy.FULL — keep everything (current default behavior, no pruning).

• MemoryStrategy.SLIDING_WINDOW — drop older tool-call exchanges, keep the most recent max_kept_turns turn-pairs in full. The LLM sees at most max_kept_turns turn-pairs on every call.

• MemoryStrategy.SUMMARIZE — like sliding_window, but instead of dropping old turns, use the LLM to generate a condensed summary. Old turns are not pruned until the summarization threshold is reached, so the LLM may temporarily see up to max_kept_turns + summarize_after_turns - 1 turn-pairs before summarization fires and compresses them back down.

Usage:

# No memory management (default — keep everything)
AgentConfig(..., memory=None)

# Sliding window with explicit parameters
AgentConfig(..., memory=MemoryConfig(
    strategy=MemoryStrategy.SLIDING_WINDOW,
    max_kept_turns=8,
))

# Summarization with explicit parameters
AgentConfig(..., memory=MemoryConfig(
    strategy=MemoryStrategy.SUMMARIZE,
    max_kept_turns=8,
    summarize_after_turns=6,
))

Parameters

strategy:

Memory management strategy.

max_kept_turns:

Maximum number of recent tool-call turn-pairs to keep in full. A turn-pair is one assistant tool-call request + its corresponding tool result(s). System prompt and user task are never counted.

Note

For strategy=MemoryStrategy.SLIDING_WINDOW, the LLM sees at most max_kept_turns turn-pairs on every call. For strategy=MemoryStrategy.SUMMARIZE, old turns are allowed to accumulate until summarize_after_turns pruneable turns exist, so the LLM may temporarily see up to max_kept_turns + summarize_after_turns - 1 turn-pairs before summarization fires.

max_tool_result_chars:

Maximum characters per individual tool result content field.

summarize_after_turns:

Only used when strategy=MemoryStrategy.SUMMARIZE. Trigger summarization when this many pruneable turns (turns outside the max_kept_turns window) have accumulated. Until the threshold is reached, old turns remain in the message list — meaning the LLM temporarily sees up to max_kept_turns + summarize_after_turns - 1 turn-pairs. Once triggered, all pruneable turns (and any prior summary) are replaced with a single summary message, bringing the count back down to max_kept_turns + 1 summary message.

summarizer_max_tool_chars:

Maximum characters to keep per tool-result message when building the summarizer input. None (default) means no truncation — the full tool result is sent to the summarizer. Set to an integer (e.g. 500) if the summarizer model has a very small context window.

summarizer_max_content_chars:

Maximum characters to keep per non-tool message (user / assistant) when building the summarizer input. None (default) means no truncation. Set to an integer (e.g. 300) for small models.

__init__(strategy: MemoryStrategy = MemoryStrategy.SLIDING_WINDOW, max_kept_turns: int = 8, max_tool_result_chars: int = 5000, summarize_after_turns: int = 6, summarizer_max_tool_chars: int | None = None, summarizer_max_content_chars: int | None = None) → None

Methods

__init__([strategy, max_kept_turns, ...])
resolve(value)	Normalize MemoryConfig | None into a validated config or None.

Attributes

max_kept_turns
max_tool_result_chars
strategy
summarize_after_turns
summarizer_max_content_chars
summarizer_max_tool_chars

strategy: MemoryStrategy = 'sliding_window'

max_kept_turns: int = 8

max_tool_result_chars: int = 5000

summarize_after_turns: int = 6

__init__(strategy: MemoryStrategy = MemoryStrategy.SLIDING_WINDOW, max_kept_turns: int = 8, max_tool_result_chars: int = 5000, summarize_after_turns: int = 6, summarizer_max_tool_chars: int | None = None, summarizer_max_content_chars: int | None = None) → None

summarizer_max_tool_chars: int | None = None

summarizer_max_content_chars: int | None = None

static resolve(value: MemoryConfig | None ) → MemoryConfig | None

Normalize MemoryConfig | None into a validated config or None.

Resolution rules:

• None → None (no memory management).

• MemoryConfig(strategy=FULL) → None (FULL means no pruning).

• MemoryConfig(...) → returned as-is.