AAperture
Back/harness engineering

Your Harness, Your Memory

Updated 2026-04-13
3 min read
597 words

Your Harness, Your Memory

Author: @hwchase17 (Harrison Chase, LangChain CEO) Core thesis: Agent harnesses and memory are deeply coupled. If you don't own your harness, you don't own your memory — and model providers are incentivized to create lock-in through memory.

Harnesses are not going away

Despite predictions that models will "absorb" scaffolding, the opposite has happened: 2023-era scaffolding (RAG chains, complex flows) has been replaced by a new type of scaffolding — the agent harness.

Examples: Claude Code, Deep Agents, Pi (OpenClaw), OpenCode, Codex, Letta Code.

Evidence: When Claude Code's source leaked, it revealed 512k lines of code — that's the harness. Even the makers of the best models invest heavily in harness engineering.

Harnesses are intimately tied to memory

"Asking to plug memory into an agent harness is like asking to plug driving into a car." — Sarah Wooders

Memory is just a form of context. The harness decides:

  • How AGENTS.md / CLAUDE.md files load into context
  • How skill metadata is presented (system prompt? system messages?)
  • Whether the agent can modify its own system instructions
  • What survives compaction and what's lost
  • Whether interactions are stored and made queryable
  • How memory metadata is presented to the agent
  • How the working directory and filesystem are exposed

At this stage in the industry, there are no well-known or common abstractions for memory. How the harness manages context and state is the foundation for agent memory.

The lock-in risk

Mildly bad: stateful APIs

OpenAI's Responses API, Anthropic's server-side compaction — storing state on their servers means you cannot swap models and resume previous threads.

Bad: closed harnesses

Claude Agent SDK (which uses Claude Code under the hood, not open source) interacts with memory in unknown ways. Any client-side artifacts it creates are non-transferable.

Worst: entire harness + long-term memory behind an API

Claude Managed Agents puts literally everything behind an API, locked into Anthropic's platform. Even open-source Codex generates an encrypted compaction summary unusable outside the OpenAI ecosystem.

Model providers are incredibly incentivized to do this because memory creates lock-in that models alone cannot.

Why memory matters

Memory is what allows agents to:

  • Get better as users interact with them
  • Build a data flywheel
  • Personalize to each user
  • Create differentiated, sticky experiences

"Without memory, your agents are easily replicable by anyone who has access to the same tools."

Chase's personal example: his internal email assistant (built on Fleet) accumulated months of memory. When it was accidentally deleted, recreating it from the same template produced a vastly worse experience — he had to reteach all preferences and tone. This made him realize how powerful and sticky memory is.

The open alternative

To own your memory, you need an open harness:

  • Open source
  • Model agnostic
  • Uses open standards (agents.md, skills)
  • Can bring your own database as memory store

This is the motivation behind Deep Agents.

Counterpoints & Gaps

  • As memory abstractions mature, standalone memory systems might make sense — but not yet.
  • The coupling argument assumes current harness architectures; future standards could decouple memory from specific harness implementations.
  • "Open" vs "closed" is a spectrum; some closed harnesses expose enough memory APIs to enable partial portability.

Sources

Linked from