Agent Memory Retrieval: Score-Space Landscape (2D)

Retrieval Score Weights

Semantic similarity w_sim 0.60

Recency w_rec 0.25

Importance w_imp 0.15

Memory capacity 30

Time speed 1.0×

Live State

Sim. time

0.0 s

Memories stored

Forgotten (evicted)

Top match score

0.00

Top-5 Retrieved

How it works

Same retrieval formula as the 3D version, drawn as two flat 2D panels instead of a rotating embedding sphere. Each memory m carries an embedding angle θ_m on a unit circle, a creation time t_m, and an importance i_m ∈ [0,1]. For a query angle θ_q, the score is:

score(m,q) = w_sim · cos(θ_q − θ_m) + w_rec · exp(−(t_now − t_m) / τ) + w_imp · i_m

cos(θ_q − θ_m)   dot product of two 2D unit vectors = cosine similarity
τ = 15 s          (recency half-life constant)

Top: embedding circle. Every memory's raw position in embedding space — its angle θ_m — plotted directly on the unit circle, colour-coded by its current score against the live query.

Bottom: score-space landscape. The same memories re-plotted with x = similarity cos(θ_q−θ_m) and y = recency exp(−age/τ). The background shading is the score formula itself evaluated at every (x,y) point (using the store's mean importance, since importance is a third axis this flat view can't show directly) — the exact plane w_sim·x + w_rec·y + w_imp·ī. A memory's true score also depends on its own importance, so a point can sit slightly off the shading it stands on.

w_sim, w_rec, w_imp — reweight the three terms live; both panels' colouring and the landscape shading update instantly.
Memory capacity — the store is bounded. New memories form automatically over time (and via "Add Memory"); once the count exceeds capacity, the single lowest-scoring memory against the current query is evicted — this is the forgetting step.
New Query — samples a fresh random query angle. Watch every point's x-position (similarity) slide instantly in the score-space panel.
Gold rings mark the current top-5 retrieved memories; point radius encodes importance, colour encodes score.

Top-5 retrieved Mid retrieval score Low retrieval score