Memory that lives in places, then teaches the rest of the mind.

Loci is not a chatbot and it is not a maze game. It is a laboratory for a particular kind of intelligence: the kind that binds a moment to a place, keeps similar places from smearing together, and later finishes a path from a fragment — a doorway, a smell of coffee, the click of an elevator.

The name is from the method of loci, the old memory-palace trick: you remember by putting things in rooms. Biology does a version of this for real. In the hippocampus, some neurons fire only when you are in a particular spot. They are called place cells. A map made of those cells is a chart of an environment. Sleep replays the day on that chart so slower parts of the brain can reuse it.

The agent here is self-improving in a narrow, checkable sense. After training, it proposes a slightly mutated child. The child is kept only if it does better on apartments it was not trained on, and is not much worse on problems of a different kind. If that gate never accepts anyone, the version number stays put. That is the point — improvement is a claim you can fail.

You do not store a photograph of your apartment. You store a walk. A handful of brain areas, mostly in the medial temporal lobe, turn that walk into something you can stand in again later.

Grid cells in entorhinal cortex fire in a triangular lattice as you move — a graph paper the rest of the system can read. Dentate gyrus granule cells fire very sparsely, so two kitchens that share a sink but not a window get different codes and do not overwrite. CA3 is an autoassociative net: give it part of a pattern, it settles to a stored whole. CA1 place cells tile the current room with Gaussian “I am here” fields; when you go upstairs those fields remap, even if the floor plan is similar. During rest, sharp-wave ripples replay recent sequences, forward and reverse, and that replay is a teaching signal for neocortex.

This is why you can lose your keys, walk the rooms once, and later finish the search from the hallway. It is also why a new office does not usually delete your kitchen, and why a nap after studying helps. Loci is a compressed, runnable cartoon of that loop — not a biophysical simulation of membranes, and not a claim that 48 units are a dentate gyrus. The names are there so the parts stay honest about what they are trying to do.

Each box is a real function in the running agent, sized so a train-and-bench pass finishes in this browser in about a second. The mapping is functional, not anatomical.

1. EC

   Entorhinal cortex

   Grid cells. A coordinate system that tiles space at a few scales.

   Nine oscillators: three spacings, three angles. Not the raw maze.

2. DG

   Dentate gyrus

   Pattern separation. Similar rooms get almost non-overlapping codes.

   Forty-eight granule cells. Only four fire (winner-take-all).

3. CA3

   CA3

   Pattern completion. A doorway retrieves the rest of the walk.

   A Hopfield attractor plus a bank of 320 engrams.

4. CA1

   CA1 place cells

   Where am I, in this room. Fields remap when the room changes.

   Sixteen Gaussians. Centers jitter per chart; sigma tightens with visits.

5. SWR

   Sharp-wave ripples

   Sleep replay. Hippocampus teaches slower cortex overnight.

   Sleep replays sequences forward and backward into the Q policy.

Functionally, most of it can. Anatomically, no — and the fly makes that obvious, because a fly does not have one “memory organ.” It has two circuits that split the job Loci currently mashed into a hippocampus cartoon.

The mushroom body is the associative memory. Projection neurons fan out onto about two thousand Kenyon cells; only a sparse set fire; dopamine neurons write valence at the output. That is the same expansion-and-sparsify motif as dentate gyrus, cerebellum, and piriform cortex. Charles Stevens called it the Marr motif. Loci’s 48-cell dentate layer is a tiny version of that, not a tiny CA1.

The central complex is the navigator. Ellipsoid-body neurons hold a bump of activity that is the fly’s heading — a ring attractor, the same idea as mammalian head-direction cells. Downstream, the fan-shaped body adds, rotates, and compares vectors: which way I am facing, which way I am moving, which way the goal is, how to steer. Flies can learn a cool spot in a hot arena, but they are not tiling a floor with Gaussian place fields. They are integrating a path on a compass.

So the honest mapping is a fork. Keep Loci’s names if you want the rat story: grids, dentate, CA3 completion, CA1 fields, ripple replay. To make it fly-like you would split the agent. Local view (the 5×5 fovea, treated like an odor) would go through a Kenyon-style expansion into a handful of output neurons, with a dopamine-like write when reward and the sparse code coincide. “Where am I” would stop being 16 Gaussians and become a heading bump plus a goal vector — then steering is “turn until they match,” which is what PFL neurons appear to do. Pattern completion of a walk from a doorway is the piece that does not port. That is a CA3 trick. The fly stores associations at the output synapse, and completes a direction, not an episode.

A “more real neural model” of the current lab, still mammalian, would also look different under the hood: spikes instead of rates; behavioral-timescale plasticity at CA3–CA3 instead of a classical Hopfield outer product; dual encodings in CA3 (sparse mossy-fiber examples and dense entorhinal concepts, which a 2024 CA3 model actually needs); theta as a clock; dopamine cells instead of a scalar surprise gate. Those are better physics. They are not required for the claim the lab tests — that a sparse expansion, an attractor, remapping charts, and replay can be reused and scored on problems you can follow. The fly is the existence proof that the expansion step is not a mammal-only story. Place cells still are.

Every tick, and every training episode, walks the same pipeline. The lab lights a stage when it is actually doing work — not a slide deck of boxes.

1. 01Task environment

   A 9×9 apartment, kitchen, hallway, or block. Walls, keys, a door, a curtain that looks solid, an elevator that opens on a beat. The agent never sees the true grid — only a 5×5 window around itself.

2. 02Sensors

   Each tick: position, last move, whether it is holding keys, a noisy local view. Adjacent tiles are legal to feel (a wall in front of your face). Distant walls stay fog until walked up to.

3. 03Perception

   The window is turned into a 64-number state: place-cell activity, grid-cell activity, local bits, last action, which problem this is. A small world model guesses the next square and the reward. Surprise is how wrong that guess was.

4. 04Hippocampus

   CA1 says where. Dentate gyrus makes the code sparse so two similar apartments do not overwrite. CA3 stores the episode and, later, completes it from a partial cue. Surprise-gated traces become engrams.

5. 05Executive

   A router picks a tactic: explore, follow a remembered route, work backward from the goal, reframe a fake wall, wait. A planner walks the occupancy map it has actually seen — unknown cells are treated as maybe-open, known walls as blocked.

6. 06Decision

   Five moves: north, east, south, west, wait. A linear Q head scores them. Legal moves that walk into a visible wall are masked. The chosen action hits the environment; prediction error updates memory.

7. 07Sleep / self-improve

   Replay restitches the day into the slower policy. Then a child agent is proposed. It is kept only if held-out apartments get better and transfer problems do not collapse. That gate is the claim of self-improvement — not a training-loss drop.

8. 08Evaluation

   Matched ablations (no place cells, no CA3, no replay…) run the same curriculum. Hippocampal benches B1–B5 and hypotheses H1–H5 are allowed to fail. A green badge is not the default.

The world is a 9×9 grid. That sounds like a toy, and it is, on purpose. A person can see the whole apartment at a glance. The agent cannot. Its eyes are a 5×5 fovea. Unvisited cells stay fog. The planner is not allowed the true maze — if it were, a hidden shortest-path search would solve every problem and we would learn nothing about memory.

Five actions: north, east, south, west, wait. Fifty-six steps is a failed episode. Reward is reaching the goal (keys, door, coffee maker, elevator, the long way around). Some problems move the goal mid-walk. Some require holding an object. One has a curtain that looks like a wall until the executive reframes it. One has a barrier that opens on a beat.

What is being tested is not “can a neural net traverse a maze.” It is whether a hippocampal-shaped memory — sparse codes, completion from a cue, remapping, replay — makes reuse and recall work on problems a person can narrate without jargon.

Each problem is a real-world wrapper around a grid family. Same engine, named so the claim is readable. The watch line is what you should look for in the live arena.

• Apartment

  Lost keys

  You dropped your keys somewhere in the apartment. Walk the rooms until you see them. After a few trips, a glimpse of the hallway should be enough to finish the path.

  Place cells light the room you are in. Sleep, then Recall — CA3 should complete the walk from a doorway.

  One-shot spatial memory. CA3 stores the episode; CA1 says where you are.

• Kitchen

  Coffee moved

  The coffee maker used to sit by the window. Someone moved it mid-morning. The rest of the kitchen is the same — only the goal jumped.

  At step 12 the goal relocates. Prediction error should spike, then the map should update without wiping the walls.

  Remapping the goal, not the whole chart. Hippocampus edits one association.

• Front hall

  Keys, then the door

  Keys are in the bowl. The front door is locked. Get the keys, then leave. Order matters.

  Subgoal flips from key to door. Working memory has to hold 'I have the keys'.

  Sequences, not single places. Replay stitches K then D into one episode.

• Office tower

  Same building, other floor

  Floor 3 is floor 2 rotated. You should reuse the first turn you already learned, not start from zero.

  Strategy should read 'replaying a remembered route'. Pattern completion pulls the sibling chart.

  Reuse. Dentate gyrus separates the floors; CA3 still retrieves the shared topology.

• Studio

  The curtain wall

  That partition looks solid. It is a curtain. Walk through it. The obvious map is wrong.

  The dashed tile is a decoy. The router has to reframe, or the planner treats it as a wall forever.

  Re-representation. The hippocampus can store 'this looked like a wall and wasn't'.

• Lobby

  Elevator wait

  The elevator doors open on a beat, then close. Rushing the door is how you miss it. Wait, then go.

  Barrier B. The agent should Wait while it is closed instead of bouncing off it.

  Time in the same place. Place cells plus a phase — not just x,y.

• Block

  Don't take the alley

  The obvious shortcut is a cul-de-sac. The long way around is the way that works. You have to notice the trap.

  A U-shaped dead end around G. Working backward from the goal beats walking in.

  The map is for planning, not for the nearest green tile.

During the day the hippocampus writes. During sleep it reads itself back to cortex. That is the biological story of systems consolidation, and it is the self-improve loop here.

Pressing Sleep does two things. First, sharp-wave-style replay: stored sequences are sampled, sometimes reversed, and written into the slower Q policy so a route can be reused without walking it again. Weak unused traces are pruned. Place-field sigma keeps the tightening it earned from visits; spatial information usually rises. Second, the agent may propose a child with slightly mutated weights. Promotion is not automatic. Validation success must rise by about five points, and transfer must not fall by more than ten. Rejected children die. The version badge in the lab is a history of that gate.

Replay is also what makes one-shot memory useful to a slow learner. CA3 can complete a walk after a single episode. The Q head cannot. Sleep is the handshake between those timescales. Turn replay off in the lab (the Replay switch) and H3 asks whether generalization actually drops. Sometimes it will not. That is a real result, not a failed demo.

B1 through B5 are claims about the circuit, not about maze scores. Each row trains a full agent and a matched lesion, then reports a number you can read.

• B1

  Pattern completion

  From a doorway, CA3 reconstructs the rest of the walk.

  Train on apartments, sleep, cue with the first third of a remembered path, score how much of the remaining path comes back. Compared to an agent with CA3 recall turned off.

• B2

  Pattern separation

  Two similar apartments get different dentate codes, so they do not overwrite.

  Same doorway in two different wall layouts. Dentate overlap across rooms should be lower than overlap of a room with itself.

• B3

  Home, then office, then home

  Learning a second neighborhood should not wipe the first.

  Train the apartment, train a rotated floor, freeze learning, retest the apartment. Charts remap so the two maps are not forced to share one code.

• B4

  One-shot recall

  A single walkthrough is enough to complete the route from a partial cue.

  One episode, then sleep, then complete from the second cell. Cortex (the Q policy) is slow; this test never trains past the first walk.

• B5

  Place fields sharpen

  Experience plus sleep tightens CA1 fields. Spatial information rises.

  Skaggs information, a standard neuroscience measure, computed from visit-weighted firing. Sigma starts wide and shrinks toward a floor as a cell is used.

Falsify retrains the full agent and six lesions on the same seed: no place cells, no replay, no CA3 recall, no strategy switching, no world model, no concept memory. Then it marks H1–H5 supported or dead. A dead hypothesis is more useful than a pretty dashboard.

• H1

  The full agent searches less than one that cannot switch tactics.

  Unique cells visited on matched maps. Lower is the prediction.

• H2

  Place cells improve success on held-out apartments.

  Validation success vs the same agent with place cells removed.

• H3

  Replay improves generalization.

  Held-out success vs an agent that never replays.

• H4

  The promotion gate reaches 60% rolling success in fewer episodes than a frozen first version.

  Episodes-to-60%. Falsified if children are never accepted, or accepted children learn slower.

• H5

  CA3 pattern completion helps transfer: a partial cue should retrieve a stored chart instead of relearning it.

  Transfer success vs no-CA3.

You can also lesion by hand. The Adjust panel has a switch per module. Train with CA3 off, then Recall — completion should collapse. Train with place cells off — the sixteen Gaussians go dark and the policy is local features only. That is the same science as Falsify, just slower, and you can watch it.

1. 01

   Pick a problem

   Start with Lost keys. It is the cleanest spatial story. The chip row under the buttons is the curriculum.

2. 02

   Train

   Forty episodes is enough to fill the engram bank and pull sigma down. Eighty or one-sixty if you want a stabler policy. The version badge may tick if a child is accepted.

3. 03

   Sleep, then Recall

   Sleep replays the day. Recall cues CA3 with the start of a recorded walk and draws the completed path. Hit rate is how much of the held-out tail came back.

4. 04

   Play or Step

   Play runs the live loop. Step is one tick: you can read the thought, the place-cell glow, the dentate winners (four of forty-eight), and the plan dots on the arena.

5. 05

   Benchmarks

   Runs B1–B5 against lesions. Takes a few seconds. Rows mark supported or falsified. Read the note under each score.

6. 06

   Falsify

   The heavier suite: H1–H5 plus full ablation reports. Use it when you want the architecture, not just the hippocampus, on the hook.

It is not a rat, and it is not a fly. Sixteen place cells and forty-eight granule cells are a schematic, sized so the math is inspectable and the benches finish while you wait. There are no spikes, no theta, no dopaminergic compartments, no body. The “apartment” is nine cells on a side. The fly mapping above is computational, not a second animal in the lab.

The policy is linear Q. The world model is a small multilayer perceptron. CA3 is a classical Hopfield net, not a modern Hopfield with softmax attention, and not a transformer memory. Surprise-gated writing is a stand-in for novelty, not dopamine. Remapping is implemented by giving each wall-pattern its own jittered chart, which is the right qualitative idea and a crude quantitative one.

Nothing leaves this browser. Maps you train stay here until you hit Reset. If a benchmark comes back falsified, believe the number. The write-up is the hypothesis; the lab is the experiment.