Memory is where a mind begins.
We build neuroplastic memory — Hebbian pathways that strengthen with use, spreading activation, decay — and measure what it does on standardized benchmarks. We measure before claiming, cite every source, and publish what didn't work.
What we've measured
These numbers came out of building and red-teaming a deployed AI agent over six months. They aren't claims — they're benchmark results, each one traceable to a recorded run.
Long-conversation memory is unsolved
Most AI memory benchmarks test single-session recall. LoCoMo tests across 10 conversations and 1,986 questions spanning multi-hop reasoning, temporal questions, and adversarial categories. Our full-ring baseline, on a routed multi-model stack:
One retrieval pipeline doesn't fit every question
Single-hop and multi-hop questions respond to different retrieval signals — what retrieves the right evidence for one starves the other. Our production retrieval stack tunes channel weights, recall depth, and reranking per category. On the conversation it was tuned on:
Hebbian pathways resist retrieval bias
Standard memory systems retrieve by similarity — and under adversarial conditions they retrieve confidently wrong answers. Memories linked by Hebbian co-activation develop lateral inhibition: semantically similar but contextually wrong memories suppress each other at recall time. This is the finding documented in our CCN 2026 paper.
Research
Neuroplastic Memory Resists Retrieval Bias: Hebbian Pathways in a Deployed AI Agent
We describe an AI agent memory system grounded in Hebbian learning and spreading activation. Memories that co-occur in useful recalls strengthen their connections; memories unused over time decay. The result is a retrieval system that learns from its own history — without retraining or fine-tuning.
How we think about this
Most AI memory work is engineering work: how to store, index, and retrieve faster. We approach it as a measurement problem first.
Empirical before architectural
We don't propose a system design until we have a number showing the existing approach fails. Every architectural decision is backed by an ablation result.
Adversarial by default
Our benchmarks include adversarial categories designed to find failure modes before deployment does. If a system can't pass adversarial recall, it isn't ready.
Long-context focus
Single-session memory is a solved problem. We focus on what happens across dozens of conversations, thousands of questions, months of use. That's where the interesting failures live.
Pre-registered gates
Ship criteria are locked before results arrive. Negative results get published alongside wins — what didn't work is data, not embarrassment.
What we build in the open
moss
Apache 2.0 · shippedOpen-source library: sanitized Hebbian memory, spreading activation, RRF retrieval, temporal disambiguation. The public core of the system behind the findings above.
github.com/Lichen-Research-Inc/moss →Hypha
Private previewA collaborative memory partner. An AI that grows hyphal pathways through your conversations — structurally coupled, long-horizon, symbiotic.
Request access →Why “Lichen”
Make an inquiry
Thirty minutes. No pitch — an honest assessment of fit. Engagement scope is determined in the consultation; we price by complexity and outcome, not templates.
Stay in the loop
Findings and paper updates. No pitch, no cadence — only when there's something real to share.
Subscribe by email →