No LLM on the write path
Memory updates are arithmetic on a graph — zero model calls per interaction.
v0.4 research previewMemory that learns
Memory that learns
what matters — for
any agent you use.
FERNme models the person, not the transcript. It runs on your own machine, speaks MCP and REST so a coding agent, a chat client, or a website can all share one profile, and updates by arithmetic with no model call on the write path — so cost stays flat as memory grows.
Apache-2.0 · 283 tests passing · runs locally · SQLite or Postgres
Flat token cost
The prompt card holds about 25 tokens on day one and after five years.
Any agent, one profile
An MCP server plus a REST API, with packaged plugins for Codex and Claude Code.
Glass-box memory
Every preference is visible, editable, and explainable through why().
Consent-gated
Recall is blocked unless the person has granted consent for that purpose.
One memory, wherever the agent runs
- MCPfernme-mcp, for any MCP-speaking agent
- Codexpackaged plugin marketplace entry
- Claude Code & Coworkpackaged plugin marketplace entry
- RESTobserve, card, recall, edit, export, delete
- FERNmarkMarkdown documents with validated tags
- Obsidiandeterministic vault import, review-queued
Local by default: memory lives in your own SQLite or Postgres store, with documents in a vault you control. Encrypted cloud sync is the project’s end goal and is not built yet.
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.