Percepta Unveils Spotlight: Decoupling Intelligence from Unbounded Memory
On October 3, startup Percepta, from Christos Tzamos' team, unveiled Spotlight, an architecture that decouples intelligence from memory: the compute module's weights stay frozen, while knowledge, skills, and working state live in a writable, unboundedly growing memory, so the model improves without retraining. The team claims it is the first design where access cost does not grow with memory size—each token reads and writes unbounded memory at constant generation cost.
Confirmed
- Spotlight replaces attention with indexable memory, described as combining attention's growing memory and capacity with the linear compute cost of fixed-state recurrent models
- It targets the core long-context trade-off: standard attention's KV cache grows quadratically, while linear-attention models have fixed capacity; Spotlight proposes a third path
- The team validated the design by embedding Python and its evolving ecosystem directly into the LLM, claiming it can host continuously evolving knowledge and surpasses attention on long contexts
Unconfirmed
- teortaxesTex questions efficiency: despite linear/constant asymptotic complexity, the constant factor may be large since one memory access touches 9 units, each involving attention computation
- Real-world performance at scale awaits independent verification
Why it matters
- If the claims hold, LLMs could accumulate knowledge and skills without weight changes or retraining, offering a path for long-running autonomous agents
- Decoupling memory capacity from compute cost challenges the attention-centric long-context mainstream, making its follow-up papers and open-source releases worth tracking
2026-10-03 ~ 2026-10-03 · 7 related posts
Primary sources
- [source] Percepta unveils Spotlight architecture: intelligence and memory separated, memory grows with no access cost — Recoil42 · 2026-10-03
- Spotlight memory architecture grows LLM capacity with linear cost, beats attention on long context — marcbhargava · 2026-10-03
- [source] Spotlight architecture adds growing memory to LLMs with constant-cost token generation — teortaxesTex · 2026-10-03
- Spotlight architecture scrutiny: attention-style memory may carry a large constant factor — teortaxesTex · 2026-10-03
- Spotlight claims a path past context caps by building Python's ecosystem inside a model — burny_tech · 2026-10-03
- Spotlight paper proposes growing memory with constant per-token cost, between Attention and linear attention — EugeneVinitsky · 2026-10-03
1 near-duplicate retellings: Recoil42