Skip to main content

​
What it is

memfmt is a small specification and a dependency-free Python library for storing an agent’s memory as plain Markdown. Mengram’s export writes it, but the format belongs to neither the product nor the company: the library works on a folder, with no account, no server and no network.
Source and specification: github.com/alibaizhanov/memfmt (MIT).

​
Three kinds of file

Agents forget in three different ways, so there are three kinds of note.

Entities

What is true. A person, a service, a file — with its facts and its relations to other entities.

Episodes

What happened, dated, and how it turned out. An event with no outcome teaches nothing.

Procedures

A workflow that was learned, with the record of it working and the log of why it was revised.

​
A procedure file

This is the one worth reading closely — it is the file the format exists for.
A workflow on its own is a guess somebody wrote down. With a record and the revisions that produced it, it is evidence — and an agent can tell a step that survived eleven deploys from one nobody has ever run.

​
Why the heading is not success / total

A bare ratio punishes the revision: a new v3 opens at 0✓/0✗ and reads worse than the v2 it was written to fix, so anything comparing the two keeps choosing the version that already failed. Progressive delivery and CI met this years ago — a canary confidence record, a flake quarantine ledger — and both smooth against a prior instead of comparing raw counts. So the percentage is a smoothed estimate, and the word after it says which kind of claim it is: The raw counts in frontmatter are exactly what happened. The percentage is derived and is never a source of truth.

​
What a step consumes and produces

The brackets are optional, and they are names rather than types. Told only that step 3 failed, you cannot tell whether step 3 is broken or whether step 1 handed it something broken — and that is the difference between fixing the right thing and rewriting a step that was never at fault. memfmt validate checks that the chain joins up and reports any step asking for something no earlier step produces.

​
Rules worth knowing

  1. Frontmatter is the source of truth. Editing the (v3 · 86% reliable) in a heading changes nothing — the parser reads the frontmatter.
  2. memfmt_type marks a file as ours. Files without it are ignored, so a memory folder can live inside a vault full of your own notes.
  3. A missing record means untracked, not failed.
  4. Round-trip or it is not the format. Parsing what was serialised returns the same object, and serialising what was parsed is byte-identical. memfmt validate checks exactly that against a real folder.

​
Where files stop being enough

Honestly: at a few hundred of them. Relevance in memfmt context is word overlap, so it misses things phrased differently; syncing a folder between machines is a real problem; deduplicating facts that contradict each other needs a model. That is what the hosted product is for. But the library stands on its own and does not expire if you never touch it — which was the condition for publishing it at all.