When conversations with AI stretch too far, the model starts to lose track — earlier messages fade from its memory as the conversation grows, and meaning can become garbled. While larger context windows go some way toward solving this problem, they merely delay that steady drift in understanding. Amp, the coding assistant from Sourcegraph, is trying to fix that by changing how its users move from one idea to the next. The company has replaced its long-standing “compaction” feature with something new: Handoff — a tool meant to carry context forward without dragging along the entire past. Amp Sourcegraph has replaced Until now, Amp gave users a way to manage long threads manually, by running a compaction (/compact) command that condensed the history into a short summary before continuing from there. It was a practical fix for a familiar problem: as conversations stretched on, the assistant would lose focus and become muddled. Compacting a thread trimmed that excess, freeing up space in the model’s limited memory window and restoring some sharpness to its replies. In effect, it let users perform a quick cleanup without abandoning the thread altogether — a reset of attention rather than intent. Handoff, according to Amp, takes a different approach. Rather than condensing the past, it creates a new thread from the old one — packaging the relevant context, setting a clear goal, and letting users review what gets passed on. The result is a sequence of linked steps, each with its own focus. How handoff works When invoked — typically by running the /handoff command in the Amp CLI or selecting Handoff from the thread menu in the interface — the feature lets you specify a goal for the new thread. Amp then analyzes the current discussion to generate a draft prompt, along with a list of files or artifacts it thinks are relevant. In the editor extensions, you can also access Handoff directly from the token usage hover, making the transition available without leaving your coding environment. /handoff /handoff now implement this for teams as well, not just individual users/handoff execute phase one of the created plan/handoff check the rest of the codebase and find other places that need this fix /handoff now implement this for teams as well, not just individual users /handoff /handoff execute phase one of the created plan /handoff /handoff check the rest of the codebase and find other places that need this fix /handoff In the example below, a user finishes mapping out an implementation plan in Amp, then uses the /handoff command to move that context into a fresh thread. The assistant analyzes the existing discussion, generates a draft prompt for the next phase, and opens a new workspace for implementation — leaving the original plan untouched. /handoff The Codex compaction concern A recent internal report from OpenAI offers a useful data point for understanding why Amp moved away from compaction. recent internal report from OpenAI Among several findings, the Codex team noted that its automated compaction system — which summarized a session and restarted it whenever the model’s context window neared its limit — was contributing to a gradual decline in performance over time, a trend that emerged soon after the arrival of GPT-5 Codex in September 2025. As sessions accumulated more “compacts,” accuracy fell, and “recursive summaries” (i.e. summaries of summaries) began to distort earlier reasoning. OpenAI’s engineers traced part of the problem to the very mechanism meant to keep conversations continuous. The report included a chart showing how the share of Codex sessions involving compaction more than doubled through October. Indeed, in the wake of OpenAI’s report publication, “Dan Mac,” a senior engineer at Amp, drew the parallel bluntly. “You should basically never use compaction,” he wrote on social media. “Thank me later.” wrote on social media One commenter framed the decision as a question of taste. Amp, they noted, is giving users control over what context moves forward through Handoff, but (notably) not which model handles it — a deliberate constraint that favors consistency and simplicity over full flexibility. It’s a trade-off between power and polish, and, as the commenter put it, “now the market decides." the commenter put it The context correction Ultimately, Amp’s move sits within a broader rethinking of how AI systems handle context. Across the industry, tools are being rebuilt to measure, extend, or externalize memory, such as the new open source Context-Bench benchmark for “agentic context engineering,” Anthropic’s agent skills layer for Claude Code, and ecosystem experiments in generating contextual scaffolding around tasks. Each tackles the same problem from a different angle: how to preserve continuity without distortion. new open source Context-Bench benchmark Anthropic’s agent skills layer for Claude Code generating contextual scaffolding around tasks Amp’s Handoff joins that trend, not by stretching context windows further, but by teaching systems — and users — how to let go of them cleanly.

Sourcegraph’s Amp Tries a New Fix for the Long-Conversation Problem

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

7 Major Learnings from The AI Engineering SF World Fair 2025

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

7 Major Learnings from The AI Engineering SF World Fair 2025

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps