FERNmePERSONALIZED MEMORYGitHub
Glass-box · built and tested

The memory is
yours to read,
correct and erase

Most agent memory is written by a model, kept by a vendor, and invisible to the person it describes. FERNme inverts all three: preferences are stored as plain weighted edges, shown in an editor, and removable on request.

  • You can read itThe glass-box editor lists every preference the engine holds about you.
  • You can question itwhy() returns the observations and outcomes behind any stored preference.
  • You can end itExport the profile, or erase it from the graph and the population prior.
Your personalized memory, arranged around youPreferences, topics, goals, context, and outcomes are held as connected lattices above a grid centred on the person. Hovering a cluster shows its current fuzzy edge weight on the engine’s 0–9 scale.
Preferences8
pref:trail-running
pref:merino
Topics7
topic:hydration
topic:ultra
Outcomes6
outcome:purchase
outcome:return
Context5
context:evening
context:mobile
Goals4
goal:replace-kit
goal:first-50k

What the engine actually holds

The memory card

What the agent actually receives: roughly 25 tokens, holding only where this person deviates from the population prior.

card · 24 tokens
  • pref:trail-running8
  • context:evening6
  • goal:replace-worn-kit5

Ask it why

Every edge can be interrogated. No black box, and no explanation generated after the fact by a model.

why(user, pref:trail-running)

11 observations · 3 good outcomes · 0 bad · last seen 4 days ago

Open in editor

Results move the weights

record_outcome() reinforces the edges behind something that worked and weakens the ones behind something that did not.

record_outcome
  • pref:trail-running 7 → 8
  • topic:road-shoes 4 → 3
  • context:evening 5 → 6

One controlled vocabulary

Every tag is canonicalised into a namespace, so the same concept does not drift into three spellings over months.

namespaces
  • pref: 23
  • topic: 18
  • goal: 12
  • context: 9
vocabulary.py

Deviations Only

The card stores where you differ from the crowd, not what you share with it.

Outcome-Weighted

What actually worked gets stronger; what backfired decays away.

Style and Mood

Tracks how a person writes and how their mood is trending, with the trend visible.

Flat Cost

About 25 tokens per turn whether the profile is a day or five years old.

Portable

Documents arrive as FERNmark Markdown or an Obsidian vault. The open protocol is spec stage.

FERNme is an Apache-2.0 research preview at v0.4, with 283 passing tests.
Published results are synthetic or LLM-authored; a real-human pilot is the pending next step.