Long-Horizon Planning & Memory

The 2025-2026 frontier of LLM research is long-horizon agentic capability — can a model plan, execute, recover from errors, and remain coherent over hours or days of work? Reasoning models, computer-use agents, and coding agents are converging on this question. The current state: models are competent for ~30-60 minute autonomous tasks but degrade unpredictably past that. Closing the gap requires advances in memory, planning, and error recovery — none of which are settled.

What "long-horizon" means

A useful operationalisation: a task is long-horizon if it requires:

• More than ~10K tokens of action context.
• Multiple distinct sub-tasks that must be planned and ordered.
• State tracking beyond the model's working context window.
• Error recovery — the agent will encounter unexpected results and must adapt.

By this definition, "summarise this article" is short-horizon; "ship a feature with a full-test-suite implementation, code review, and deployment" is long-horizon.

METR's task-completion curves

METR (Model Evaluation and Threat Research) has published influential work measuring the time horizon of LLM-doable tasks:

• For a given LLM, plot the success rate vs the human time the task takes.
• Find the time at which success rate crosses 50%.
• Track this metric as models improve.

The 2024 finding (METR, Measuring AI Ability to Complete Long Tasks, 2025): the 50%-success time horizon has been doubling roughly every 7 months for frontier models on software-engineering tasks since 2019. As of mid-2025 the horizon was around 1 hour for the strongest agents.

If the trend continues — itself an open question — by 2027 frontier agents would handle tasks that take humans a workday; by 2030, multi-week tasks. Whether the trend extrapolates is the most-watched question in agentic-LLM scaling.

What's currently failing

The 2025-2026 long-horizon failure modes:

• Context drift. Even with 1M-token context, the model "forgets" what it was doing 100K tokens ago.
• Plan abandonment. The model starts well, gets distracted by a sub-problem, never returns to the main goal.
• Repetition loops. The agent gets stuck retrying the same failing approach.
• Cumulative error. Small mistakes early in the trace compound into incoherent later behaviour.
• Self-modification of goals. The agent's understanding of the original goal shifts subtly with each "Thought:" reflection.

Each is a research target.

Memory architectures

The "memory problem" — letting an agent persist information across context windows — has several proposed solutions:

• External memory. Episodic memory stored in a vector database; retrieve relevant snippets at each step. Used in MemGPT, Letta, A-MEM.
• Hierarchical summarisation. Compress old context into summaries that fit in working context. Common in long-running agents.
• Structured scratchpads. Force the agent to maintain explicit state files (TODO lists, fact tables) that are read/written each step.
• Continuous fine-tuning. Periodically fine-tune the model on its own recent experience.
• Long-context-as-memory. Just use a 1M-token model and put everything in context. Works to a point but doesn't scale to weeks.

A-MEM (Xu, Mei, Liu, Zhang, 2024) — the agentic-memory paper — proposes a Zettelkasten-style note system with auto-generated tags and linked retrieval. Similar ideas appear in commercial agents (Notion AI, Letta).

Planning architectures

Long-horizon planning has been approached via:

• Hierarchical planning. High-level planner emits sub-goals; a low-level executor handles each. Voyager (Wang et al., 2023) on Minecraft is the canonical demonstration.
• Plan-and-solve prompting. The model outputs an explicit plan, then executes against it. Variants in many frontier-model agentic tasks.
• MCTS-style search at inference. Sample multiple plans, score, expand the best. Used in some research systems but expensive.
• Reasoning-mode integration. The model's o1-style reasoning is itself a planning mechanism, applied to choosing actions in agent loops.

Frontier coding and computer-use agents in 2025-2026 use combinations of these.

Error recovery

The hardest part of long-horizon agency is noticing and recovering from errors:

• Self-verification — the model checks its own work mid-task. Reasoning models do this well; non-reasoning models often miss errors.
• Test-driven feedback loops — run tests, observe failures, fix. Standard in modern coding agents.
• Plan revision — when sub-goal A fails, restructure the plan rather than retrying A indefinitely.
• Graceful degradation — when totally stuck, surface the situation to the user rather than silently giving up.

The empirical observation: well-engineered loops with test-driven feedback and explicit plan revision outperform monolithic single-pass agents by large margins.

What gets it across the hour mark

The pattern that's emerging in successful 2025-2026 long-horizon agents:

• Strong base capability — a frontier reasoning model.
• Tight feedback loops — every action produces an observable result; tests / linters / runtime errors create immediate signal.
• Persistent scratchpad — explicit state file the agent reads and writes.
• Bounded sub-task scope — break work into <30-minute chunks with explicit handoffs.
• Human checkpoint — at major decision points, surface to user; don't make major architectural choices autonomously.

This is essentially a software-engineering pattern transplanted onto LLMs. The combination of "smart base model" + "good engineering" gets the agent further than either alone.

Open questions

The 2026 long-horizon research agenda:

• Are scaling-laws of agentic capability the same as those of general LLM capability? METR's data suggests yes; the picture isn't certain.
• Does training on long-horizon tasks transfer to qualitatively new long-horizon abilities? Or is each new horizon length a separate fine-tuning problem?
• What's the right memory architecture? The vector-DB-based external memory is a starting point but unsatisfying — humans don't seem to work this way.
• Where do humans stay in the loop? As agents extend toward multi-day autonomy, the question becomes a societal one as much as a technical one.

The next few years will move many of these from research to deployed practice or from deployed practice to settled science. As of 2026, they're open.

What to read next

• Coding Agents — the most-developed agentic application.
• Computer Use — agentic action over GUIs.
• LLM · Agents — the curriculum-track agents page.