Harness, Memory, Context Fragments, & the Bitter Lesson
Author: Viv Trivedy (@Vtrivedy10) Source: X thread, 2026-04-12 A work-in-progress mental dump on the intersection of harness design, memory accumulation over long timescales, and the search bitter lesson we can't escape.
Harnesses & Context Fragments
A crucial job of the harness is to efficiently and correctly route data within its boundaries into the context window for computation.
The context window is a precious artifact. Harnesses make explicit decisions on how to populate, manage, edit, and organize it. Each loaded object can be thought of as a Context Fragment — an explicit decision by the user and harness designer about what the model needs to do work at any given time.
Many ideas on externalizing objects and loading them into context are pioneered by @a1zhang with RLMs (Reinforcement Learning Machines).
Experiential Memory
We're in the very early days of deploying agents, and agents produce massive amounts of data in every interaction. This is analogous to humans doing things and remembering them.
Agent memory has a massive advantage: it can be accumulated across all agents, which are easily forked and duplicated (unlike humans). As @dwarkesh_sp discusses, this is a massive benefit of artificial systems.
Memory can be treated as an externalized object. The harness is tasked with contextualized retrieval — pulling in the right data from accumulated memories across all agent interactions.
Search & The Bitter Lesson
As we deploy agents over year-long timescales, there will be a hyper-exponential growth in data produced by those agents. We should want to:
- Own that data — open ecosystems are critical
- Use that data — which means searching over, distilling, and organizing massive amounts of it
Our brains are exceptional at this: using prior experience contextually, and committing the right stuff to memory through intentional practice. Our current infrastructure and algorithms will be put to the test and often break in this new data regime.
Open questions
- How do we efficiently distill experiences (traces) into higher-level memory primitives that capture the important parts — over ultra-long time horizons?
- How much of the future is search just-in-time vs search integrated into model weights?
- How do we make models much better at self-managing their context window?
- How do we reduce error rates in recursively allowing agents to operate over external objects?
Related
- harness-engineering/your-harness-your-memory — Harrison Chase on harness-memory coupling
- harness-engineering/overview — Harness engineering overview
- Agent Memory Engineering(Nic) — 三种生产架构深度审计
- claude-code/llm-wiki-pattern — LLM Wiki Pattern (externalized memory objects)
- harness-engineering/memory-as-trajectories — Memory as trajectories
2026-10-10 补充:Thariq 的 harness 快速过时论
来源:Latent Space 访谈 Thariq Shihipar
Thariq 从模型换代角度补充了本页的 bitter lesson:harness 和 memory 对象一样会被模型进步快速折旧。为某个模型的 failure mode 写的规则(如 per-model CLAUDE.md)在新模型上常常 over-constrain 或完全失效,维护成本随模型代数累积。他的实践是默认不写、按重复失败模式补规则,与本页"own that data / use that data"的搜索立场形成对照:Thariq 押注的是模型吸收能力的 scaling,Viv Trivedy 押注的是外部记忆与搜索的基础设施。两个赌注的胜负取决于模型换代速度与个人数据资产复利速度的相对关系。
Counterpoints & Gaps
- "不写规则"在 Anthropic 内部可行(官方 harness 迭代快),对个人用户,模型切换的滞后意味着规则仍有窗口期价值。
- 单一访谈陈述,无 harness 折旧率的量化数据。