Marty Sampson.aiAI, applied
Explainer · source-backed

Context is not memory.

Keeping information and using the right information are two different jobs.

An assistant can appear to remember something simply because that detail is still in the conversation it can see. What happens after a fresh session is the more useful question.

An original M-brand explainer

From stored information to an answer

  1. 01 · Retain

    History + notes

    Saved exchanges and selected decisions live outside the current context.

  2. 02 · Select

    Retrieval

    Bring relevant saved material into this turn. A useful item can be missed.

  3. 03 · Assemble

    Current context

    Instructions, recent messages and selected evidence are available together.

  4. 04 · Respond

    Answer + review

    Check the response against the evidence, not just its confidence.

Compaction is a separate operation. A long working context becomes a shorter handoff. A summary may omit details; a system retaining originals can retrieve them later.
Read left to right, or top to bottom on a small screen. This is a conceptual model, not a tested installation. Durable-note writes are separately permissioned; an answer does not automatically become a saved fact.

Five jobs, not one feature

Context is material available for this turn. History records earlier exchanges. Durable notes preserve selected information. Retrieval selects what to bring back. Compaction shortens working material. Storage alone does not establish successful recall.

A detail that exists—but gets missed

Invented example, not a test result: someone plans a community-garden flyer. They choose Saturday and specify that the wording must not imply plants are free. A dated note records both decisions.

In a new session, the assistant is asked for a headline. If the note is retrieved, both constraints are available. If retrieval misses it, the history can remain stored while the headline misses a requirement. The diagnostic is: was the fact saved, selected, and used?

Now change the event to Sunday. Does the system distinguish the current decision from the older one? An evaluation needs corrections and dates, not just successful recall of unchanged facts.

What the sources establish

Context engineering

Anthropic's context-engineering article distinguishes compaction from structured notes persisted outside context and brought back later. It informs the vocabulary here; the example and visual arrangement are original.

Hermes

The Hermes memory guide documents bounded persistent notes and a separate mechanism for searching previous conversations. These are related capabilities, not interchangeable descriptions of one store.

Lossless Claw

The Lossless Claw README describes preserving raw messages, assembling summaries with recent messages, and retrieving original details. My interpretation: retained originals make recovery possible; they do not guarantee every response retrieves or uses the right detail.

Documentation rechecked. These are documented mechanisms, not local verification of performance.

A small test worth doing next

Use synthetic notes with known answers, a distractor, a changed decision and one question the notes cannot answer. Record the material supplied to the model and its response. Compare an ordinary session, a fresh session and a compacted handoff where supported.

Score correct answers, unsupported assertions, missed constraints and appropriate uncertainty separately. No memory-system experiment has been executed for this article; a reproducible MartyBytes companion is planned.

Download the evaluation checklist (.txt)

Next: What should an AI agent remember? →

Back to field notes