Back to browse
GitHub Repository

Three-path memory for LLM agents: KV window (exact attention) + retrieval index (archived chunks) + TRN recurrent state (8-96 KB, O(1) per step)

0 starsPython

3-path agent memory – 8 KB recurrent state vs. 156 MB KV at 10K tokens

by tenpa0000·Mar 14, 2026·1 point·0 comments

AI Analysis

●●●BangerWizardryBig Brain

Keeps agent memory at 8 KB constant size while KV caches bloat to 156 MB.

Strengths
  • O(1) memory complexity per step prevents latency degradation at long contexts.
  • Three-path mixing allows exact attention plus compressed patterns.
  • Benchmarks show flat throughput at 10k tokens where transformers slow down.
Weaknesses
  • Requires custom model architecture, not a drop-in library for existing LLMs.
  • Recurrent state might lose some nuance compared to full attention.
Category
Target Audience

LLM developers, Agent builders

Similar To

RWKV · RetNet · MemGPT

Similar Projects

AI/ML●●●Banger

Stateful Inference with 99% Token Savings

Injects raw KV tensors directly into model cache to skip 90% of token recomputation.

Big BrainBold Bet
wasnaga
201mo ago
Developer Tools●●Solid

Agents Remember – Git-aware memory for coding agents

Git-verified memory ledger beats vague agent context files.

Big BrainNiche Gem
foxfire_1st
3211d ago