Applying Memory to Bots

How the AI Agent can learn one opponent within a single heads-up game: the ideas behind it, what we can measure, the player types it looks for, and how we’ll know it works.

Since then: this is a plan. The baseline game, with memory switched off, comes first. The results will be the next field note, “The Bot Is Watching How I Play Now.”

Where this fits

Our AI Agent plays the same way against everyone. That is how the strongest poker programs work: Pluribus, the bot that beat elite professionals at six-player no-limit in 2019, “plays a fixed strategy that does not adapt to the observed tendencies of the opponents” [5]. A fixed strategy is hard to exploit. But it also leaves money on the table against players with obvious leaks, such as a player who bluffs every river or one who calls everything.

Memory is the step from playing well to playing well against this person. The groundwork is already in place: the agent can accept a short read on its opponent, but only at a two-player heads-up table. What’s missing is the part that watches: counting what the opponent does and turning it into a short, honest read.

Five ideas behind opponent memory

1. Count, then compare with normal

Opponent modeling starts simply: count how often a player does something each time they have the chance, and compare that with how a sound player would behave. Online players do this with a HUD (heads-up display), which shows stats like VPIP (how often a player puts money in before the flop) beside each opponent [1][2]. Research bots did the same in 1998: Loki, from the University of Alberta, kept a set of weights over each opponent’s possible hands and re-weighted them after every action it saw [3]. The version that modeled each opponent specifically clearly outplayed a version without that modeling in its self-play tests.

2. A few hands tell you much less than they seem to

This matters most for a 100-hand memory. If a player has bluffed 3 of the 5 river bets you’ve seen, the true rate could be anywhere from about 23% to 88% (a Wilson score interval). Guides disagree on how many hands a stat needs. One coaching guide says VPIP is usable after about 20 hands, flop continuation-bet stats after about 100, and river stats only after 1,000 or more [1]. A statistical calculator is stricter: about 1,000 hands for VPIP within ±2.5%, and 500 to 1,000 for three-bet stats [2]. Both points hold. A rough read is possible early; a precise one takes a long time.

Game 990605 gives a sense of the budget. In 52 hands Kevin saw 49 flops, 38 turns and 36 rivers, and 23 hands reached showdown. In 100 hands that is roughly 95 flops, 70 rivers and 40 showdowns: enough for broad reads, not fine ones.

3. Exploiting has a cost

Moving away from sound play to punish a leak also opens holes of your own. If the read is wrong, or the opponent changes style, the exploiting bot loses more than a steady one would. Research names this tradeoff. Restricted Nash Response [6] builds counter-strategies that exploit a model of the opponent only as far as a chosen confidence allows, because pure best responses “are brittle.” Safe opponent exploitation [7] goes further: only risk, through exploitation, the chips the opponent has already given away through mistakes. Their earlier DBBR approach [8] played sound poker for a while, then switched to exploiting.

For HoldemRobots this means the memory agent should start from its normal play, adjust only when the evidence is clear, adjust by degrees, and say so when there’s no clear read yet.

4. The bot may only remember what a player could see

HoldemRobots is honest by construction: the dealer shows each seat only what a fair player can see. Memory must follow the same rule.

• Cards from a fold win are not public. When a hand ends because someone folded, the winner’s cards are never shown at a real table. The memory may only use cards from hands that went to showdown.
• The table program can see everything. It holds every seat’s cards. The memory must read only the betting history and showdown results, never the hidden cards.

One asymmetry to know about: the heads-up table shows the AI’s cards to you at the end of every hand, even when nobody showed down. That’s a helpful review feature, but it means you learn more about the bot than it will learn about you.

5. Language models need facts computed for them, and memory changes how they play

A recent study of LLM poker play found that the models misjudge hand strength and opponent ranges, and that their stated reasoning often doesn’t match their actions. Giving them tools that compute the hard parts helped a lot [9]. That matches what we found in Part 10: hand facts worked out in code fixed the misreads. A 2026 preprint gave LLM poker agents persistent memory and found opponent modeling appeared only when memory was present, with each agent’s reads written in plain language that people could check [10]. Separately, work on social-deduction games found that separate, self-correcting models of each opponent improved LLM agents’ win rates [11].

The lesson for us: compute the stats in code, give the model a short plain-language read, and log that read so people can audit it.

What we can measure

Every metric below can be computed from what the dealer already records: each action with seat, street and amount, plus the cards shown at showdown. “One game?” says whether 100 heads-up hands give enough chances for a usable read.

Chances per 100 hands are rough estimates for heads-up, scaled from game 990605, and depend on both players’ styles. Stat definitions follow common HUD usage [1][2][4].

The player types memory looks for

Players are commonly sorted into a few styles, each with a known counter [12][13]. Published thresholds are for six-player or full tables; heads-up players are much looser, so the numbers below are starting guesses to tune, not rules. The memory’s job is to decide which type the opponent most resembles, how sure it is, and pass along the matching counter.

The Solid type matters as much as the others: a good memory must correctly say “no clear read” against a sensible player, and not invent a leak.

Here is the kind of line the AI would receive, 40 hands into a game:

Opponent read (40 hands): looks like a Bluffer, fairly sure. Bets when checked to
18 of 22 (82%); river bets shown down: 4 of 6 were weak. Counter: call down lighter,
check strong hands and let them bet, don't bluff them.

Early in a game, the same line says “no clear read yet (9 hands),” and the AI plays exactly as it does today.

Where it can go

The first three phases are in reach. The rest show where opponent memory can go, drawing on the research, even if we can’t get there yet.

How we’ll know it works

• Chips won per 100 hands, memory against no memory, on the same deals. Against each exploitable test bot the memory agent should win clearly more; against the Solid bot it should do about the same.
• Hands until the read is right, for each type. A good first target: the correct label within 30 to 40 hands.
• False reads: how often it labels the Solid bot as something else. This should be rare.
• Behavior change: its bluff rate should fall against the Station, and its calls should rise against the Bluffer.
• Against people: Kevin’s Part 1 and Part 2 games, then the rematch with the brother-in-law. Does it call lighter by his second or third bluff, and fold more once his river bets turn out strong?

Sources

Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. This is the public version of the Sep 26 plan; research gathered Sep 26, 2026.