TAU-HOME.COM
LOADING

Victor Taelin Introduces Unii: An Near-Infinite Context AI Harness Powered by OptMem Compaction

Victor Taelin revealed Unii (UniiChat), a personalized AI harness that maintains persistent conversations by instantly compacting contexts exceeding 128k down t

tau · October 9, 2026

#AI-Harness #Infinite-Context #OptMem #UniiChat #VictorTaelin

Victor Taelin Introduces Unii: An Near-Infinite Context AI Harness Powered by OptMem Compaction

Victor Taelin (@VictorTaelin), creator of the HVM/Bend programming language and founder of Higher-Order Company, introduced Unii (UniiChat) on October 8, 2026. Described as the closest experience to infinite context, Unii is a custom-built AI agent harness designed to maintain persistent conversational workflows through continuous real-time memory compaction.

Visualization of Victor Taelin's Unii AI harness interface and real-time context compaction pipeline

Image source: @VictorTaelin via X (Twitter)

Unii manages user conversations and tasks through an append-only log architecture. When the active context approaches model constraints, the system automatically triggers a compaction pass to keep the working context fixed. Taelin stated that he has relied exclusively on Unii over the past week to organize his daily workflow and build software projects, noting plans to release the codebase as open source next week.

Real-Time Context Compaction via OptMem

The core differentiator in Unii lies in its real-time context management powered by Taelin's OptMem memory optimization algorithm.

Conventional LLM coding and workflow harnesses often suffer from degrading attention or surging token costs when conversations expand beyond default context windows. Unii counters this by enforcing a bounded memory envelope:

  • Instant Compaction at 128k: As soon as accumulated context crosses 128k tokens, Unii immediately compacts the working state back down to approximately 64k tokens.
  • Cache Efficiency & Fixed Budget: Maintaining context around ~64k stabilizes prefix caching performance while allowing compaction passes to run on cost-effective models like Claude Haiku.
  • Append-Only Database Log: All user interactions and state transitions are stored as an immutable log, from which key facts and ongoing state are summarized into active memory.

Pure Vibe-Coding Implementation and Architecture

The development of Unii was also an experiment in end-to-end agentic synthesis. Taelin revealed that Unii was entirely vibe-coded within itself over a matter of days without manual line-by-line programming.

"I read a total of 0 lines," Taelin noted, demonstrating a self-bootstrapping process where the harness generated its own tools and UI logic through iterative natural-language prompts.

Regarding the technical stack, Unii is not currently implemented in Bend due to existing limitations with web execution and WebSocket support. Taelin stated plans to port Unii to the Bend ecosystem later this year as its runtime capabilities mature.

API Direct Invocations and Operational Caveats

Unlike subscription-based frontends subsidized by platform providers, Unii operates on direct API invocations:

  • API Usage Costs: Sustained high-token sessions and frequent compaction calls can lead to significant API bills. Taelin noted that ongoing work is focused on resolving prompt caching bugs and reducing token consumption.
  • Early Stage Bugs: Given its zero-manual-code vibe-coding origin, Taelin cautioned early users to expect bugs and interface quirks.
  • Self-Hosting and Replication: Ahead of the official open-source release, Taelin published implementation prompts and algorithm references for developers wishing to implement OptMem-style compaction in their own harnesses.

Sources