Agent Tape System

Definition

An agent tape system is an append-only record of messages, tool calls, observations, state changes, and other task events, paired with a movable working view that determines which parts the agent currently sees.

Current Synthesis

The DeepChat and Bub cases treat tape as a context architecture rather than a synonym for chat history. Preserving an ordered event record lets an agent revisit earlier evidence during long work, while a pointer, window, anchor, or handoff controls what enters the active model context. DeepChat combines the record with selective compression and inspection; Bub emphasizes hiding earlier spans without deleting them and searching the original history when needed. Both approaches can reduce information loss relative to replacing the full history with one summary, but neither eliminates retrieval, salience, cost, or privacy problems.

Tape belongs inside the wider Agent Harness: the record only becomes useful when the runtime can decide what to retain, expose, summarize, inspect, or replay. Its inspectability also makes it useful for diagnosing tool calls, reasoning content, multi-turn interaction, and failed loops.

Key Claims

• Append-only storage preserves task history while allowing the active context to move independently.
• Agents can revisit earlier records during long tasks instead of relying entirely on lossy context compression.
• Tape inspection improves debugging because model requests, tool calls, observations, and loops remain traceable.
• A tape is not automatically good memory; selection, compression, retrieval, and relevance still determine what the model can use.
• Persistent records require privacy, retention, and access controls when they contain local files, credentials, or enterprise work context.
• Anchors and handoffs can express task-stage boundaries without requiring destructive compression of earlier records.
• Historical search is itself an agent behavior that needs query discipline and selective loading to avoid flooding active context.

Evidence

Long-task context continuity

• VOL.002｜DeepChat：为什么要做一块开源 AI 试验田？ says DeepChat combined append-only context, a position pointer, and compression into a Tape-based harness that lets agents review earlier records.

Inspection and debugging

• VOL.002｜DeepChat：为什么要做一块开源 AI 试验田？ describes Trace and Tape inspectors used to examine tool calls, reasoning content, multi-turn behavior, and model loops.

Anchors, handoffs, and selective retrieval

• Ep 58. 从零实现龙虾需要什么？Bub 开发者访谈 describes Bub recording conversations, tool calls, state, intermediate events, and feedback, then using anchors and handoffs to hide earlier spans while retaining them for later TapeSearch-style retrieval.
• Ep 58. 从零实现龙虾需要什么？Bub 开发者访谈 also reports a failure boundary: broad historical searches can fill the context, so the agent must make room, retrieve selectively, and repeat narrower searches when needed.

Counterevidence & Qualifications

• The source does not provide controlled comparisons showing that Tape outperforms other context-management approaches.
• An append-only record can grow expensive, surface stale information, or retain sensitive material unless the harness applies retrieval, compaction, and governance.
• Tape, Persistent Agent Memory, and AI Skills overlap in practice but serve different functions: event history, durable experience, and reusable procedure should not be collapsed into one layer.
• Bub’s preference for preserving original history does not prove that avoiding compression is always more accurate, economical, private, or operationally reliable.

What Changed

• Added Bub’s anchor-and-handoff implementation as a second tape architecture.
• Distinguished nondestructive history retention from active-context visibility.
• Added selective historical search and context flooding as explicit operational boundaries.

Related Concepts

• Agent Harness - governs how tape records are written, selected, compressed, and used.
• Persistent Agent Memory - converts selected experience into durable context beyond one task history.
• Context Engineering - designs what information enters the active model context.
• AI Verification - can use retained traces as evidence when checking agent behavior.
• DeepChat - project case maintaining its own Tape implementation.
• Bub - project case using anchors, handoffs, and selective tape retrieval.
• Group-Chat Agent Participation / 群聊 Agent 参与机制 - social setting whose high message volume makes selective context necessary.