Temporal knowledge graphs · No. 4

The Memory
Observatory

A team's conversation, read the way a query-time memory reads it. Time wraps the dial; the question is the needle.

Building memory

Messages arrive around the dial. Entities are born at their first mention and sink toward the centre as they gather relations. At the end of the dial a question fires.

Demo conversation — synthetic, narratively anchored

—
0%

How to read the instrument

Three encodings from the Narrative Observatory, and one new one.

Time is the dial

Four months of a five-person team's chat run clockwise from the top, one revolution from 1 January to 30 April. Every message is a grain on the rim. The inner ring is the pulse: messages per day.

Connection is gravity

Each entity the conversation names is a star, placed at the angle of its first mention. Relations pull it toward the centre. This star graph is what a write-time memory system maintains: every message pays an extraction call, every correction a merge.

The needle is the query

When the dial completes, a question fires. The instrument replays a query-time retrieval: lexical seeds scatter across the dial, time clustering finds the dense interval, the hands swing to frame it, one labelling pass keeps the messages that matter and links them, and the forest is drawn into the answer.

The cost is the point

Nothing was paid when the messages arrived. The structure is built once, at the moment of the question, over the window the clock found. Toggle the write-time graph to see what the alternative keeps lit all the time.

What the replay follows

The four stages are the method of Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory (T-Mem, arXiv 2610.12874), computed here on the demo data without a model:

  1. Seeds. BM25 over the raw messages for the query's terms; the top fourteen hits. Some land in the right week, some are strays from January and April.
  2. Episode. DBSCAN over the seeds' timestamps. The densest interval becomes the window; the strays fall out as noise.
  3. Forest. Every message in the window is labelled Key, Support or Irrelevant (in T-Mem an LLM does this; here the labels are annotated), then linked: an Update link where a later message changes an earlier one, Expand links where context supports a key message.
  4. Answer. The kept messages are serialised within an 8,192-token budget and cited by id.

Demo conversation · synthetic and narratively anchored · previous instrument: The Narrative Observatory