01
Zero-LLM hot path
Routine memory writes are arithmetic graph updates, not extraction calls. Memory formation stays bounded as traffic grows.
Fuzzy-Edged Recall Network
Memory that gets stronger without getting expensive.
FERNme is a Hebbian-first memory layer for site agents. It updates a sparse fuzzy graph by default and reserves semantic reasoning for uncertainty.
click purchase booking message correction outcome
FERN graph
HEBBIANDECAYCONFIDENCERECALL
agent prompt site action recommendation support routing memory editor
01 / SYSTEM
Memory as infrastructure, not prompt bloat.
Routine learning stays deterministic. Semantic reasoning appears only where its cost can change the outcome.
02
Uncertainty-gated semantics
When confidence is low, FERNme can ask a small LLM or the user. Otherwise it continues learning deterministically.
03
Token-flat recall
Recall compiles the strongest relevant edges into a compact card instead of replaying a long conversation history.
04
User-visible memory
People can inspect, correct, export, or delete memory so personalization does not become hidden surveillance.
02 / METHOD
How FERNme remembers.
A small set of interpretable operations turns behavior into usable recall.
01
Observe behavior
Capture consented events such as bookings, corrections, purchases, and successful outcomes.
02
Update fuzzy edges
Strengthen or decay graph edges between signals, preferences, contexts, and actions.
03
Score confidence
Combine evidence strength, conflict, taxonomy match, recency, and outcome feedback.
04
Recall a compact card
Retrieve only the strongest relevant memories for the agent’s current task.
05
Escalate when needed
Free text, causal ambiguity, or low confidence can trigger optional enrichment or a user check.
Confidence gate
Reasoning is a fallback, not the hot path.
Learn from repeated evidence, retain uncertainty, and escalate only when an ambiguous update is important enough to justify the cost.
confidence ≥ 0.85 → act on strong memory
0.45 – 0.85 → observe more evidence
confidence < 0.45 → optional LLM or user check
prompt card → top relevant edges only
03 / RESEARCH STATUS
A working engine with an honest boundary.
WORKING
Fuzzy graph updates, compact recall, consent gates, audit history, persistence, export, and deletion.
ACTIVE WORK
Turning messy real-world site events and free text into reliable, taxonomy-rich memory signals.
POSITION
A research preview for bounded transactional memory—not a claim of general, production-ready agent memory.
04 / FAQ
Questions, answered plainly.
Is FERNme pure Hebbian?
No. It is Hebbian-first. The default path is deterministic; optional semantic enrichment is reserved for uncertain cases.
Does it require LLM calls?
No. Structured behavioral events can update the fuzzy graph without an LLM call.
How does it stay cost-bounded?
FERNme avoids always-on extraction and full-history prompting. Recall compiles a small, relevant memory card.
What does the user control?
Consent, human-readable recall, correction, export, and deletion are part of the service boundary.
Is it production-ready?
Not yet. FERNme is a research preview. The graph, recall, consent, audit, and storage paths exist; real-world ingestion and production hardening remain active work.