Memory Strategies

Short-term memory strategies

Short-term strategies control which messages are included in the LLM prompt.

MemoryMaxTokensStrategy

class MemoryMaxTokensStrategy(ShortTermMemoryStrategyBase):
    memory_strategy_id: Literal["max_tokens"] = "max_tokens"
    memory_max_tokens: int = 10000
    conditions: list[str] = []

Algorithm: Walks the conversation backwards, grouping messages by human turn (Interaction). Keeps interactions until memory_max_tokens is reached.

Field	Type	Default	Description
memory_strategy_id	"max_tokens"	required	Strategy identifier
memory_max_tokens	int	10000	Token budget. -1 = include nothing
conditions	string[]	[]	simpleeval expressions to filter interactions

JSON example:

{
  "memory_strategy_id": "max_tokens",
  "memory_max_tokens": 8000,
  "conditions": ["message.role != 'tool'"]
}

---

Long-term memory strategies

Long-term strategies trigger when the conversation exceeds a threshold and either summarize or erase old context.

Shared trigger fields

All long-term strategies inherit these fields:

Field	Type	Default	Description
input_tokens_threshold	int	0	Trigger when LLM input tokens exceed this value (0 = disabled)
interaction_keep	int	0	Interactions to keep after the strategy runs
interaction_conditions	string[]	[]	Filter which interactions count toward thresholds
interaction_threshold_qty	int	0	Trigger after this many interactions (0 = disabled)
interaction_threshold_tokens	int	0	Trigger after this many interaction tokens (0 = disabled)
on_tools_success_threshold	string[]	[]	Trigger when any of these tool names succeed
on_tools_error_threshold	string[]	[]	Trigger when any of these tool names error

SummarizerStrategy

Uses Chain of Density (CoD) summarization to compress conversation history.

class SummarizerStrategy(LongTermMemoryStrategyBase):
    long_term_memory_id: Literal["summarizer"] = "summarizer"
    interaction_threshold_qty: int = 20
    interaction_threshold_tokens: int = 20000
    cod_max_loops: int = 5
    knowledge_base: KnowledgeBases | None = None
    instructions: str | None = None
    text_splitter: TextSplitterConfig = Field(default_factory=TextSplitterConfig)

Algorithm:

1. Check if any trigger condition is met
2. Run cod_max_loops iterations of CoD summarization
3. Optionally store the summary in a vector store (knowledge_base)
4. Keep the last interaction_keep interactions

Field	Type	Default	Description
long_term_memory_id	"summarizer"	required	Strategy identifier
interaction_threshold_qty	int	20	Trigger after N interactions
interaction_threshold_tokens	int	20000	Trigger after N tokens
cod_max_loops	int	5	CoD densification passes
interaction_keep	int	0	Interactions to keep after summarization
instructions	string	null	Custom summarization instructions
knowledge_base	KnowledgeBase	null	Store summaries in a vector store
text_splitter	TextSplitterConfig	default	Text splitting config for vector store ingestion

JSON example:

{
  "long_term_memory_id": "summarizer",
  "interaction_threshold_qty": 15,
  "interaction_threshold_tokens": 15000,
  "cod_max_loops": 3,
  "interaction_keep": 2,
  "instructions": "Focus on decisions made and action items.",
  "knowledge_base": {
    "collection_id": "session-memory",
    "search_mode": "always",
    "top_k": 3
  }
}

NeuralyzerStrategy

Erases conversation history beyond a threshold. Faster than summarization — use when historical context is not needed.

class NeuralyzerStrategy(LongTermMemoryStrategyBase):
    long_term_memory_id: Literal["neuralyzer"] = "neuralyzer"
    interaction_threshold_qty: int = -1
    interaction_threshold_tokens: int = -1

Field	Type	Default	Description
long_term_memory_id	"neuralyzer"	required	Strategy identifier
interaction_threshold_qty	int	-1	Trigger after N interactions (-1 = never)
interaction_threshold_tokens	int	-1	Trigger after N tokens (-1 = never)
interaction_keep	int	0	Interactions to keep after erasure

JSON example:

{
  "long_term_memory_id": "neuralyzer",
  "interaction_threshold_qty": 30,
  "interaction_keep": 5
}

---

The Interaction model

Both strategies group messages into Interaction objects — one per human turn (human message + all subsequent AI/tool messages until the next human message).

This grouping ensures that related messages (a tool call and its result) are always kept or discarded together.

---

Related pages

• Memory concepts — when and why to use each strategy
• Agent Configuration — where to configure memory