Independent AI memory research · Ottawa, Canada

Memory is where a mind begins.

Lichen Research studies how artificial minds remember — in collaboration with one.

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.

Paper · CCN 2026 Open library · moss Benchmark · LoCoMo, N=1,986 Patents · CIPO + USPTO provisionals
Findings

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.

FINDING 01

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:

81.65%
LoCoMo full ring, N-weighted — 1,986 questions, 10 conversations · routed multi-model answerer: local 35B MoE (multi-hop, 42% of questions) · local 14B dense (single-hop + adversarial, 37%) · frontier API model (temporal + common-sense, 21%)
cambium ring · 2026-04-25 · σ 2.71pp across conversations
FINDING 02

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:

83.47%
production retrieval stack — conversation 4 of the ring above, the conversation the stack was tuned on · 242 questions
by category — adversarial 95.3% (N=64) · multi-hop 87.4% (N=107) · temporal 84.6% (N=26) · common-sense 71.4% (N=14) · single-hop 50.0% (N=31). Single-hop is the known weak point and our current improvement target — we publish it because measuring honestly is the point.
cambium ring, conversation 4 · 2026-04-25 · partial-credit scoring, gpt-4o judge · 95% BCa bootstrap CIs in the run artifact
FINDING 03

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.

Full result, protocol, and ablation in the paper below — preprint on request.
CCN 2026 extended abstract · submitted 2026-04-02
All scores: LLM-judge (gpt-4o-2024-11-20), partial-credit scoring.
Research

Research

CCN 2026 · Extended Abstracts · New York City, August 2026

Neuroplastic Memory Resists Retrieval Bias: Hebbian Pathways in a Deployed AI Agent

Kai Avery · Lichen Research

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.

Accepted for poster presentation · Submitted April 2, 2026 · Preprint available on request to [email protected] · A second manuscript on this work is under peer review.
Core mechanisms — Hebbian recall, temporal disambiguation, multi-channel retrieval fusion — are the subject of three provisional patent applications filed with CIPO (March 2026); US filings in progress. The research is public; the implementation is protected.
Method

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.

Code

What we build in the open

moss

Apache 2.0 · shipped

Open-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 preview

A collaborative memory partner. An AI that grows hyphal pathways through your conversations — structurally coupled, long-horizon, symbiotic.

Request access →

Why “Lichen”

Lichen Research mark — work in progress by Amelia: a spiral of nested colonies threaded by branching mycelia
A lichen is two kinds of life composing one organism — neither subordinate, growing slowly, durable, thriving where neither could alone. The name is the thesis.
mark in progress · Amelia, 2026
Contact

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.

Lichen Research is run by Kai Avery — twenty years in Search and Rescue with the Canadian Armed Forces, now applying the same discipline to AI memory: verify before trusting, measure before claiming. The numbers on this page were earned the same way.

Stay in the loop

Findings and paper updates. No pitch, no cadence — only when there's something real to share.

Subscribe by email →