Memory

Alquimia has two memory layers: short-term (what the LLM sees in the current prompt) and long-term (what gets summarized or erased when context grows too large).

Short-term memory

Short-term memory controls which messages from the conversation history are included in the LLM prompt. Without a strategy, all messages are included — which can exceed the model’s context window.

max_tokens strategy

Walks the conversation backwards, grouping messages by human turn (Interaction), and keeps messages until the token budget is exhausted.

{
  "short_term_memory_strategy": [
    {
      "memory_strategy_id": "max_tokens",
      "memory_max_tokens": 8000
    }
  ]
}

Parameter	Type	Default	Description
memory_strategy_id	"max_tokens"	required	Strategy identifier
memory_max_tokens	int	10000	Maximum tokens to include. -1 = include nothing
conditions	string[]	[]	Empathy-style conditions to filter which interactions to include

Long-term memory

Long-term memory strategies trigger when the conversation grows beyond a threshold. They either summarize the history or erase it, keeping only the most recent interactions.

Trigger conditions

All long-term strategies share these trigger fields:

Field	Description
input_tokens_threshold	Trigger when the LLM’s input token count exceeds this value
interaction_threshold_qty	Trigger when the number of human turns exceeds this value
interaction_threshold_tokens	Trigger when total interaction tokens exceed this value
on_tools_success_threshold	Trigger when any of these tool names succeed
on_tools_error_threshold	Trigger when any of these tool names error
interaction_keep	Number of most-recent interactions to keep after the strategy runs

summarizer strategy

Uses Chain of Density (CoD) summarization: iteratively compresses the conversation into a progressively denser summary over cod_max_loops iterations.

{
  "long_term_memory_strategy": [
    {
      "long_term_memory_id": "summarizer",
      "interaction_threshold_qty": 20,
      "interaction_threshold_tokens": 20000,
      "cod_max_loops": 5,
      "interaction_keep": 3,
      "instructions": "Focus on action items and decisions made.",
      "knowledge_base": {
        "collection_id": "session-summaries",
        "search_mode": "always"
      }
    }
  ]
}

Parameter	Type	Default	Description
long_term_memory_id	"summarizer"	required	Strategy identifier
interaction_threshold_qty	int	20	Trigger after this many interactions
interaction_threshold_tokens	int	20000	Trigger after this many tokens
cod_max_loops	int	5	Number of 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 for RAG

neuralyzer strategy

Erases memory beyond a threshold. Simpler and faster than summarization — use when you don’t need to retain historical context.

{
  "long_term_memory_strategy": [
    {
      "long_term_memory_id": "neuralyzer",
      "interaction_threshold_qty": 30,
      "interaction_keep": 5
    }
  ]
}

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

Combining strategies

You can combine short-term and long-term strategies:

{
  "profile": {
    "short_term_memory_strategy": [
      { "memory_strategy_id": "max_tokens", "memory_max_tokens": 8000 }
    ],
    "long_term_memory_strategy": [
      {
        "long_term_memory_id": "summarizer",
        "interaction_threshold_qty": 20,
        "cod_max_loops": 3,
        "interaction_keep": 2
      }
    ]
  }
}

Flow: When the long-term strategy triggers, it summarizes the conversation and keeps the last 2 interactions. The short-term strategy then limits what the LLM sees to 8,000 tokens from those kept interactions.

Persistence strategies

After each inference, the conversation is persisted according to persistence_strategy:

Value	Behavior
INCREMENTAL	Append new messages to the existing session (default)
FLUSH	Replace the session with the current conversation
EPHEMERAL	Do not persist — session is lost after inference

Related pages

• Memory Strategies reference — full API reference
• Agent Configuration — where to configure memory