aglet-builtin-memory-summary 0.1.0a1

Aglet memory technique: rolling LLM-summarised history for long conversations.

aglet-builtin-memory-summary

Rolling LLM-summarised conversation memory — the right Memory Technique for long chats where a sliding window would drop important context.

Part of the Aglet pluggable Agent runtime.

What it does

When the per-conversation buffer grows past trigger_chars, the technique asks an LLM to compress the oldest half of the history into a single short paragraph and recalls that summary on every subsequent turn. Recent turns are kept verbatim.

Best combined with memory.sliding_window under routing: parallel_merge:

memory:
  techniques:
    - { name: sliding_window, config: { max_messages: 10 } }
    - name: summary
      config:
        model: cheap
        trigger_chars: 6000
        keep_recent: 6
  routing: parallel_merge

Install

pip install --pre aglet-builtin-memory-summary

Config

Key	Default	Description
model	default	Model alias (from agent.yaml models:) used for compression
trigger_chars	6000	Compress when the conversation buffer exceeds this many chars
keep_recent	6	Number of most-recent messages kept verbatim
summary_prefix	"[Prior conversation summary]"	Label prepended to recalled summaries

License

Apache-2.0