Today I’m releasing a new experiment, LOLA.

Lola is an autonomous AI agent with long and short term memory, a twitter account, and a trading wallet loaded with 5 sol.

She can scan new pairs, analyze token metrics, and trade autonomously, with reflections on her actions stored in long term memory to refer back upon later.

Her core personality takes on an optimistic, naive trader- one who’s fresh on chain and sees the potential in every coin no matter how dumb. As she makes trades, she can get “smarter”- referring back on her reflections stored in long term memory and social interactions.

Lola’s Architecture

Lola’s architecture is composed of several modular components:

Core / Memory System

• Langchain for memory, ReAct for reasoning.

• Long term memory takes on a unified “messy” approach, in which technicals/ narrative/social memories are not silo’d, in attempt to mirror the way we all experience trading.

• After trades, Lola records a few reflection sentences on her decision and outcome.

• Core personality, writing, and reflections use a fine tune of Mistral’s Large model, which I’ve found to be wonderful at capturing a consistent character and tone across outputs.

• Data analysis & computer vision use Claude 3.5 Sonnet as Lola’s “quant”, which allows her core personality to be less clinical and more natural.

Trading System

Lola’s trading system loops thru this cycle:

• Pulls new tokens from data API and filters out likely rugs basic metric (holder count, age, market cap). Checks for recycled twitter accounts, etc.

• Tokens that pass initial check are shared with Lola, where she can make a decision to buy or pass, based on her sentiment towards the name, ticker, description, and on chain metrics. Initially, she can purchase between 0.25-1 sol of a token.

• Once position is opened, Lola will check token’s current price, holder count increase/decrease, and chart on each trading cycle.

• No set logic for when she must sell- data is simply passed to her, and response is formatted as action (hold, sell X %, or sell all), along with reflection on choice.

• Lola can also choose to buy tokens stored in short term memory, via tweets replying to her containing cashtags.

Basically- the hope is to mirror the way most of us trade- check telegram, check new pairs, look at open positions, repeat.

Social System

• Her posts are derived from her core personality, trading reflections, and recently queried replies/ mentions on X.

• As her long term memory fills with experiences, I expect personality will begin to shift- probably a bit more cynical after being rugged a few times lol.

Narrative Explorer

• Lola queries to top 100 new tokens every once a day, and attempts to find common themes or narratives to signal emerging metas. Most recently noticed trend is stored in short term memory to refer back upon. In theory, she’ll be able to start gaining a grasp on recurring themes & trends.

WIP & Future Modules

There’s tons left to implement and optimize, but here’s some of the features that will roll out for Lola as soon as i hack them together:

• Hack together TikTok trending API to get latest top searches / add to memory and weigh heavy if a token matches later.

• Trending News data feeds: allowing her more context and the ability to recognize emerging stories in the case of something like $PNUT.

• Multi worker underlings who can research each trending coin more in depth and report back summaries to Lola.

• Deep search- tracing funding sources and connections across top holders of a coin for overlap- augmented by AI allows for more detailed analysis than just statistical. ( for ex, pointing out that xyz addresses were all funded by CEXs on same day)

• Telegram chat from unified “single perspective”- I.e. mirroring the experience of one person bouncing between groupchats.

Token?

We all know the way these things work. We tokenize everything- it doesn’t really matter if I wanted it or not. Within a few hours there would be a token bearing her name and image, attributed to me.

SO- when first started up, Lola has a “run once” pipeline, in which she can choose a ticker (she’ll prob choose $LOLA tho because it’s in her training data) and then fire deployment tx programmatically.

I’ll Dexscreener update for her- my treat.

-echo