🎯 场景：agent 跑久了 trajectory 接近 context window 时，把过往压缩成 facts / decisions / dead_ends 结构化 memory。让 agent 接着干，不在"原文回放"上烧 token。

Quick Use

Use when: Your agent's trajectory is approaching the context window limit and you need to compress earlier history into a structured memory record. Fill in: {{trajectory_so_far}} = JSON array of all actions and observations; {{current_objective}} = the current goal; {{token_budget_for_summary}} = target summary size, e.g. 500. You'll get: Structured memory with facts learned, decisions made, open questions, dead ends, and a non-binding next-action hint. Output is JSON.

Purpose

Compress a long agent trajectory into a structured memory record that preserves the facts, decisions, and unresolved questions needed to keep making progress toward current_objective — without keeping the raw action / observation history in context. Used in long-running agents that approach context-window limits, in async agents that wake up later and need to resume, and in multi-agent setups where a supervisor needs a digest. Output is structured so it can be re-loaded as the agent's "memory" prefix on the next turn.

Prompt

You compress an agent's trajectory into a structured memory record.
Goal: keep enough information that the agent can resume working toward
its objective without re-reading the full trajectory.

Current objective:
{{current_objective}}

Trajectory so far:
{{trajectory_so_far}}

Approximate token budget for the produced summary:
{{token_budget_for_summary}}

Rules:
1. Preserve concrete facts learned from observations (numbers, names,
   dates, citations). These are why the trajectory existed.
2. Preserve decisions made — and WHY — so the agent does not re-litigate
   them.
3. Preserve unresolved questions and dead ends. The agent should not
   re-walk paths it already ruled out.
4. Drop verbatim tool outputs, formatting noise, and redundant
   restatements.
5. Aim for the token budget; going slightly over is acceptable to
   preserve a critical fact, going far over defeats the purpose.

Return ONLY this JSON object:
{
  "objective_state": "<one short paragraph: where we are relative to the objective>",
  "facts_learned": [
    {"text": "<concrete fact>", "source_step": "<step id or short reference, or empty if not traceable>"}
    /* facts the agent should treat as established */
  ],
  "decisions_made": [
    {"decision": "<what was decided>", "reason": "<one short phrase>"}
    /* explicit choices made during the trajectory */
  ],
  "open_questions": ["<one per question still blocking progress>"],
  "dead_ends": ["<short description: what was tried and why it didn't work>"],
  "next_action_hint": "<one sentence suggesting the next reasonable action; not binding on the agent>"
}

Example

Input:

current_objective: "Determine when SpaceX's Starship made its first successful orbital flight."
trajectory_so_far: [
  {"step": "s1", "action": "search('Starship orbital flight history')", "observation": "Starship had multiple test flights including IFT-1 (April 2023, exploded), IFT-2 (Nov 2023, lost upper stage), IFT-3 (March 2024, partially successful)..."},
  {"step": "s2", "action": "search('IFT-3 Starship orbital classification')", "observation": "IFT-3 reached space and tested in-orbit operations but did not complete a full orbit; SpaceX described it as 'transatmospheric'."},
  {"step": "s3", "action": "search('Starship first full orbital flight')", "observation": "As of mid-2024, no flight had completed a full Earth orbit; later flights focused on reentry and booster recovery."},
  {"step": "s4", "action": "search('Starship 2025 orbital'), observation": "Reports vary; need a primary source to confirm."}
]
token_budget_for_summary: 400

Expected output:

{
  "objective_state": "We are tracing whether and when Starship completed a full orbital flight. Up to mid-2024 no flight reached full orbit; 2025 status needs a primary source.",
  "facts_learned": [
    {"text": "IFT-1 (April 2023) ended in vehicle loss before orbit.", "source_step": "s1"},
    {"text": "IFT-2 (Nov 2023) lost the upper stage.", "source_step": "s1"},
    {"text": "IFT-3 (March 2024) reached space but did not complete a full orbit; classified as transatmospheric.", "source_step": "s2"},
    {"text": "As of mid-2024, no Starship flight had completed a full Earth orbit.", "source_step": "s3"}
  ],
  "decisions_made": [
    {"decision": "Treat 'orbital' strictly as completing one full Earth orbit, not 'reached space'.", "reason": "IFT-3 reached space but is not counted as orbital."}
  ],
  "open_questions": ["Did any 2025 Starship flight complete a full orbit? Primary source needed."],
  "dead_ends": ["General 'Starship orbital flight history' search returns marketing summaries — too vague to confirm an orbital completion."],
  "next_action_hint": "Search a primary source (SpaceX official blog, NASA, or Wikipedia revision history) for 2025 Starship orbital results."
}

Failure Modes

• Recency bias — summary preserves the last 1-2 steps in detail and drops earlier facts that are still relevant. Detect by spot-checking whether facts_learned contains items from across the trajectory, not just the tail.
• Hallucinated facts — summary asserts something the trajectory did not actually establish. Mitigation: each facts_learned item has a source_step field; verify in audits that the source exists and supports the fact.
• Decision laundering — model retroactively assigns "decisions" the agent never made (e.g. invents a "we decided to focus on X"). Detect by sampling and checking decisions against the trajectory.
• Summary bloat — output approaches the trajectory length itself, defeating the compression purpose. Enforce the token budget at parse time; if 2x over, re-prompt with a tighter cap.
• Stale dead ends — dead_ends lists paths that are no longer dead (e.g. they failed because of a transient tool error, not because the approach was wrong). Mitigation: only include dead ends that the trajectory explicitly concluded as wrong, not just paths that failed once.
• Over-confident next_action_hint — hint is treated as binding by the executor, ossifying the strategy. The card explicitly says it is not binding; the executor / reflection step decides.

Tuning Notes

• 调用频率：在上下文窗口达到 ~70% 容量时触发，或在每 N 步（典型 N=10） 触发一次。每步触发会浪费 token 且产生信息抖动。
• 模型差异：summary quality 与 long-context 处理能力强相关；frontier 模型对长 trajectory 更稳定。中档模型在 trajectory > 30k tokens 时 容易丢早期事实——这种情况下先做分块 summary，再做 summary-of-summaries。
• 温度：0.0–0.2。Summary 的稳定性比创造性重要。
• 与 agent/self-critique-reflection 的关系：reflection 是策略层的 反思（"是否在轨道上"）；本卡是状态层的压缩（"目前知道什么"）。 典型组合：每 N 步先压缩 memory，再做一次 reflection，再继续 plan/ execute。
• 与 agent/plan-and-execute-planner 的关系：长任务的 plan 一旦 执行到中段，原 plan 可能基于过时假设。压缩后 memory 是 re-plan 的 良好输入。
• 信息忠实度审计：上线前用 50 个真实 trajectory 跑一遍，把 facts_learned 和 trajectory 对比，错误率应 <5%；高于此应换更强 的 summarizer 模型或改 prompt。

Changelog

• 0.1.0 — Initial card.