Skip to main content

Class: ObservationCompressor

Defined in: packages/agentos/src/memory/observation/ObservationCompressor.ts:122

LLM-based compressor that takes a batch of ObservationNote objects and produces denser CompressedObservation summaries.

Achieves 3-10x compression while preserving key facts, entities, and temporal context. Each compressed observation carries three-date temporal metadata: when the compression happened, the earliest referenced event, and a human-friendly relative time label.

Constructors​

Constructor​

new ObservationCompressor(llmInvoker, traits?): ObservationCompressor

Defined in: packages/agentos/src/memory/observation/ObservationCompressor.ts:127

Parameters​

llmInvoker​

(system, user) => Promise<string>

Function that calls an LLM with (system, user) prompts.

traits?​

HexacoTraits

Optional HEXACO personality traits for bias-aware compression.

Returns​

ObservationCompressor

Methods​

compress()​

compress(notes): Promise<CompressedObservation[]>

Defined in: packages/agentos/src/memory/observation/ObservationCompressor.ts:144

Compress a batch of observation notes into denser summaries.

The method:

  1. Formats the notes as a numbered list for the LLM.
  2. Sends the batch to the LLM with a compression prompt.
  3. Parses the JSON array response into CompressedObservation objects.
  4. Attaches three-date temporal metadata (observedAt, referencedAt, relativeLabel).

Parameters​

notes​

ObservationNote[]

Batch of observation notes to compress.

Returns​

Promise<CompressedObservation[]>

Array of compressed observations. Returns empty array on LLM failure.