---
title: "Momentum · concept · Agentic Atlas"
description: "Momentum: how context already in the window steers the model along its current path."
canonical: "https://agentic-atlas.dev/nodes/momentum"
last-updated: "2026-09-23"
---

1. [Agentic Atlas](https://agentic-atlas.dev/)
2. [Patterns in the Agentic Atlas](https://agentic-atlas.dev/atlas)
3. [Context Flow](https://agentic-atlas.dev/nodes/context-flow)
4. Momentum

Seen working in [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window).

1. concept
   # Momentum
   **Why can earlier context keep steering an agent in the wrong direction?**
   Resident context conditions each next token, so a path already in the window biases its own continuation.
   Hook
   how context already in the window steers the model along its current path
   Laws & fences
   - Waiting never breaks a failure loop, because each new attempt adds more of the same pattern.
   - Mid-conversation corrections mute an old path, but only removing it from the input makes it absent.
   - Direction set early costs least because steering cost grows with trajectory, yet momentum is a tendency, not a lock.
   - Reset when correcting the current path costs more than starting over.
   When to reach
   - Reach for this when an agent keeps retrying variants of the same failed approach.
   - Reach for this when a mid-conversation correction looks complete but the old direction may persist.
   - Reach for this when deciding whether drift comes from resident context or from the model's training.
   Provenance
   [momentum/definition](https://agentic-atlas.dev/nodes/momentum#definition) · v1.0.11
   Addresses
   atlas_cards momentum
2. [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window)
3. [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment)
4. [Context Flow](https://agentic-atlas.dev/nodes/context-flow)
5. [LLM Dispatch](https://agentic-atlas.dev/nodes/llm-dispatch)
6. [Statelessness](https://agentic-atlas.dev/nodes/statelessness)
7. [The Context Economy](https://agentic-atlas.dev/nodes/context-economy)
8. [Context Efficiency](https://agentic-atlas.dev/nodes/context-efficiency)
9. field notes
   - “validate at the return seam; re-dispatch over repair; bounded autonomy — the contract keystone, applied”
   - “state the handle and purpose immediately before a governed act — momentum's good pole deliberately seeded, plus a legible receipt”

Context Efficiency unfold the map fold the map

The card, in place · its connections drawn edges from atlas_links momentum

On this plate

[definition](https://agentic-atlas.dev/nodes/momentum#definition) [why-it-matters-the-problem](https://agentic-atlas.dev/nodes/momentum#why-it-matters-the-problem) [model-and-claims](https://agentic-atlas.dev/nodes/momentum#model-and-claims) [scope-and-boundaries-three-fences-and-a-test](https://agentic-atlas.dev/nodes/momentum#scope-and-boundaries-three-fences-and-a-test) [implications-the-consequence-map-eviction-s-second-motive](https://agentic-atlas.dev/nodes/momentum#implications-the-consequence-map-eviction-s-second-motive) [evidence-the-grounding-pass](https://agentic-atlas.dev/nodes/momentum#evidence-the-grounding-pass) [relationships](https://agentic-atlas.dev/nodes/momentum#relationships) [lineage](https://agentic-atlas.dev/nodes/momentum#lineage) [open-questions-todo](https://agentic-atlas.dev/nodes/momentum#open-questions-todo) [relationships-ledger](https://agentic-atlas.dev/nodes/momentum#relationships-ledger)

Every section is addressable on its own. Read only the ground you need.

## Definition

[Permalink to Definition section](https://agentic-atlas.dev/nodes/momentum#definition)

**The window steers, not just informs.** The substrate is autoregressive conditioning: resident context shapes the next-token distribution, so a trajectory biases its own continuation. One studied candidate mechanism on the beneficial side is the **induction head** — an attention circuit implementing `[A][B] … [A] → [B]`: find the earlier occurrence, copy what followed. Olsson et al. (2022) report strong causal evidence in small attention-only models and correlational evidence in larger models; induction heads are not established as the universal mechanism of in-context learning.

## Why it matters — the problem

[Permalink to Why it matters — the problem section](https://agentic-atlas.dev/nodes/momentum#why-it-matters-the-problem)

Designers reason about resident context as information: is it correct, is it relevant, what does it cost. But resident context also *conditions the trajectory*: once an approach, a tone, a formatting choice — or a mistake — is in the window, generation is biased toward continuing it. An agent that has failed the same way three times tends to try variants of the same failure, precisely because the failures are resident. No relevancy or correctness accounting captures this: perfectly accurate context can still trap the actor on a path. And the trap is easy to misread — mid-window course corrections *look* complete while the old heading is still in every forward pass.

## Model and claims

[Permalink to Model and claims section](https://agentic-atlas.dev/nodes/momentum#model-and-claims)

- **One force, two directions.** In-context learning and few-shot prompting are momentum pointed well — resident demonstrations select the task and the model continues it (Brown et al. 2020; Xie et al. 2021 frames this as implicit Bayesian inference over a latent concept). The failure modes are the same force pointed wrong: hallucination **snowballing** — the model over-commits to an early wrong claim and defends it, even when it can recognize the error in isolation (Zhang et al. 2023) — and the repetition loop, where each repetition *raises* the phrase's probability, a literal positive feedback loop (Holtzman et al. 2020). No established research term unifies the two poles; the vocabulary is pole-specific (ICL vs. exposure bias / snowballing / degeneration). *Momentum* names the one force the literature describes twice. (Unity holds at the substrate level; whether the *same circuit* produces both poles is contested — assert conditioning, not circuitry.)
- **Momentum compounds by self-reseeding, not by aging.** Individual tokens lose grip with distance — a system-prompt persona measurably drifts within ~8 rounds (Li et al. 2024), and the *last* items in the window steer the next output disproportionately (recency bias, Zhao et al. 2021). But a live trajectory never lets its pattern get old: every continuation deposits fresh evidence of the pattern at the recency-privileged tail. Three failed attempts pull hard not because they have sat there long, but because the newest tokens are always more of the same failure. Corollary: **waiting never breaks a doom loop; only removal does.**
- **Decay mutes; only eviction removes.** A stateless model re-reads the whole transcript every turn, so everything resident is *in the calculation* every turn, decayed or not. In-window course correction re-weights the old heading; it cannot remove it — early anchors keep biasing output even under explicit "ignore that" instructions (Lou & Sun 2024), and a dormant pattern can re-fire when its trigger recurs. Muted, not gone. The economy re-bills every resident token; momentum re-weighs it — every turn, pay and pull. Only an input-changing boundary can subtract that trajectory; admitting more context can counterweight it, but cannot make the resident path absent.
- **Turn radius.** Steering cost scales with the mass of resident trajectory — a freighter's turns are measured in miles. Set the course while the window is light (structural priming persists measurably: Sinclair et al. 2022); past some mass, correction costs more miles than a restart. Yet momentum is a tendency, not a lock: a single pointed challenge can collapse the model's own prior answer entirely (Xie et al. 2024) — entrenchment is strongest when *unchallenged*.
- **[Placement](https://agentic-atlas.dev/glossary/placement) modulates it.** The pull is not uniform across the window: attention privileges the head (attention sinks — a structural artifact of causal attention: Xiao et al. 2024) and the tail (recency), and sags in the middle (Liu et al., *Lost in the Middle*, TACL 2024). Head and tail carry momentum; the middle is where it goes to die.

## Scope and boundaries — three fences and a test

[Permalink to Scope and boundaries — three fences and a test section](https://agentic-atlas.dev/nodes/momentum#scope-and-boundaries-three-fences-and-a-test)

*(Concept — a dynamics phenomenon of the pipeline, observed, not applied.)* Nothing here prescribes a design step: momentum is a phenomenon, not a move.

What momentum is not:

- **Not friction.** [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment)'s friction is contradictory content resident together — a property of a *frozen snapshot*, always negative signal. Momentum is a property of the *trajectory over time*, sign-neutral. Different axes; the same transcript can carry both readings.
- **Not the prior.** Steering that survives a reset was never momentum — it is the model's weights (training distribution, RLHF): gravity, a field acting on every trajectory, even at rest. A reset zeroes momentum; it cannot zero gravity. Boundary case: a *standing admission* (CLAUDE.md, a system prompt loaded every session) behaves like a prior across instances but is momentum mechanically — it steers because it is resident, and you can edit it.
- **Not sycophancy.** Bending toward the *user's* stated view is a distinct, user-directed force; momentum is the pull of the model's *own* trajectory (Pan et al. 2025 formalizes the axis as user bias vs. assistant bias; Sharma et al. 2023 for sycophancy proper).
- **The identity test — the fresh-dispatch counterfactual.** Would a clean window, given the same task, still drift that way? *Yes* → prior, not momentum. *No* → momentum. A fresh dispatch isn't just 100% [relevancy](https://agentic-atlas.dev/glossary/relevancy); it is **zero momentum** — which is why it doubles as the test.

## Implications — the consequence map: eviction's second motive

[Permalink to Implications — the consequence map: eviction's second motive section](https://agentic-atlas.dev/nodes/momentum#implications-the-consequence-map-eviction-s-second-motive)

Momentum across the lifecycle: **[admission](https://agentic-atlas.dev/glossary/admission) seeds it** — priming is cheapest while the window is light; the rudder moved before the ship has speed sets the whole course. **[Residency](https://agentic-atlas.dev/glossary/residency) compounds it** — by self-reseeding (above). **[Eviction](https://agentic-atlas.dev/glossary/eviction) resets it** — which gives eviction and fresh dispatch a **second motive** beyond the economy's budget arithmetic, in two modes:

- **Emergency reset** — break a doom loop. In Laban et al. (2025), instructions revealed across simulated multi-turn conversations produced a **39% average performance drop** across 15 models and six analytical generation tasks; the same shards concatenated into one fresh prompt averaged **95.1% of the fully specified control**. That 95.1% is a single-turn control, not a post-failure recovery rate. The paper's two-model Recap arm improved on the sharded condition but remained below the Full and Concat controls. The bounded result supports *consolidate before retrying*; it does not measure a universal restart dividend.
- **Waypoint reset** — *planned* seams that make large course changes affordable. A turn too big for one window is segmented like a linear approximation of a curve: run a leg, distill the endpoint, dispatch fresh, run the next leg. Each seam is a free heading change. Reflexion is the canonical instance: discard the failed trajectory, carry only the distilled lesson forward (Shinn et al. 2023).

**Summarization is a *partial* reset** — a summary carries the conclusions but sheds the trajectory (compaction in practice: Anthropic 2025). Course correction concentrates at the seams; mid-trajectory it costs miles.

And every reset that breaks bad momentum also discards good momentum: the fresh window that escapes the doom loop forgets the conventions the trajectory had usefully established. The partial reset (summary, envelope) is the compromise position, and the turn radius gives the decision rule: **reset when turning costs more than restarting.**

## Evidence — the grounding pass

[Permalink to Evidence — the grounding pass section](https://agentic-atlas.dev/nodes/momentum#evidence-the-grounding-pass)

Verified 2026-07-04 (three-agent research pass; URLs fetched). Anchors dated because they rot.

- **Good pole / mechanism:** Brown et al. 2020, *Language Models are Few-Shot Learners* (NeurIPS, arXiv:2005.14165); Xie et al. 2021, *ICL as Implicit Bayesian Inference* (ICLR 2022, arXiv:2111.02080); Olsson et al. 2022, *In-context Learning and Induction Heads* (Transformer Circuits, arXiv:2209.11895).
- **Bad pole:** Zhang et al. 2023, *How Language Model Hallucinations Can Snowball* (ICML 2024, arXiv:2305.13534); Holtzman et al. 2020, *The Curious Case of Neural Text Degeneration* (ICLR, arXiv:1904.09751); lineage: exposure bias (Bengio et al. 2015, NIPS; Ranzato et al. 2016, ICLR) — *contested*: He et al. 2021 (EMNLP) finds distortion limited with a "self-recovery" capacity, so state compounding as tendency, not law. No umbrella term spans the poles (verdict from targeted search, 2026-07-04); circuit-level unification is live disagreement (one unreviewed 2026 preprint for; arXiv:2507.07810 against; the strongest pro paper withdrawn).
- **Reset:** Laban et al. 2025, *LLMs Get Lost in Multi-Turn Conversation* (arXiv:2505.06120) — 39% average multi-turn drop across the studied tasks; the fresh Concat control averages 95.1% of Full, while the post-trajectory Recap arm recovers only part of the gap; Shinn et al. 2023, *Reflexion* (NeurIPS, arXiv:2303.11366); ERGO (arXiv:2510.14077) — entropy-spike-triggered resets, +56.6%.
- **Decay / anchoring / priming:** Li et al. 2024, *Instruction (In)Stability* (COLM, arXiv:2402.10962); Lou & Sun 2024, *Anchoring Bias in LLMs* (arXiv:2412.06593); Sinclair et al. 2022, *Structural Persistence* (TACL); Xie et al. 2024, *Ask Again, Then Fail* (ACL, arXiv:2310.02174).
- **Fences:** Pan et al. 2025, *User-Assistant Bias* (arXiv:2508.15815); Sharma et al. 2023, *Towards Understanding Sycophancy* (ICLR 2024).
- **Placement:** Liu et al., *Lost in the Middle* (TACL 2024); Xiao et al. 2024, *Attention Sinks* (ICLR, arXiv:2309.17453); Zhao et al. 2021, *Calibrate Before Use* (ICML, arXiv:2102.09690).
- **Guard:** *context rot* (Chroma 2025) is attention dilution from long/noisy input — a different phenomenon; it motivates the economy's decay curve, not momentum's bad pole. Do not conflate.

[Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window) reads Laban's four conditions as two panels: Full → Sharded shows the force (same information, a 39% average drop once delivered as a trajectory instead of one prompt), and Recap vs. Concat isolates residency (both hear every shard; the window that never got lost wins). The dual-panel form keeps the 95.1% Concat control from being quoted as a recovery rate.

## Relationships

[Permalink to Relationships section](https://agentic-atlas.dev/nodes/momentum#relationships)

- **Child of [Context Flow](https://agentic-atlas.dev/nodes/context-flow)** — a property of the pipeline's dynamics: what is resident shapes the future of the stream.
- Gives **eviction** (flow) and **fresh dispatch** ([LLM Dispatch](https://agentic-atlas.dev/nodes/llm-dispatch)) a second motive beyond budget arithmetic — and *validate at the return seam; re-dispatch over repair; bounded autonomy — the contract keystone, applied* its dynamics rationale: discarding beats repairing because repair fights momentum.
- Complements [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment)'s friction pole: friction misleads (content), momentum entrenches (trajectory) — two distinct ways resident context hurts beyond token cost.
- Parallels [The Context Economy](https://agentic-atlas.dev/nodes/context-economy): statelessness re-bills every resident token (price) and re-weighs it (force).

## Lineage

[Permalink to Lineage section](https://agentic-atlas.dev/nodes/momentum#lineage)

Positive feedback — the deviation-amplifying loop Maruyama named as the counterpart to cybernetics' deviation-counteracting one (*The Second Cybernetics*, American Scientist, 1963): output re-entering as input reinforces whatever direction a system already holds, one mechanism behind both the useful lock and the runaway. Momentum is that loop closed by construction — counter-steering only adds a term; only removal damps.

## Open questions / TODO

[Permalink to Open questions / TODO section](https://agentic-atlas.dev/nodes/momentum#open-questions-todo)

- **No node yet manages the force as a whole.** Two moves fall out of the model, and they are the legs an unminted *momentum management* node would carry: *trajectory priming* — exploit it, seeding conventions while the window is light — and *waypoint segmentation* — evade it, chaining partial resets through a turn too big for one window. *state the handle and purpose immediately before a governed act — momentum's good pole deliberately seeded, plus a legible receipt* is the first minted technique on the priming leg, and it governs a single act rather than the leg; segmentation has no node at all. [LLM Dispatch](https://agentic-atlas.dev/nodes/llm-dispatch) and *validate at the return seam; re-dispatch over repair; bounded autonomy — the contract keystone, applied* each exercise the force for their own purpose — neither owns the steering decision.

The relationships ledger

Evidence-bearing references

## Relationships

Every connection keeps the section where it was found. The map above orients; this ledger carries the evidence.

### Outbound references 10

1. in-slice · occurrence 1
   [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment)
   how well the available context fits the task at hand
   Evidence: [Scope and boundaries](https://agentic-atlas.dev/nodes/momentum#scope-and-boundaries-three-fences-and-a-test) · occurrence 1
2. in-slice · occurrence 1
   [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window)
   why a fresh prompt beats restating requirements inside a conversation gone wrong
   Evidence: [Evidence](https://agentic-atlas.dev/nodes/momentum#evidence-the-grounding-pass) · occurrence 1
3. in-slice · occurrence 1
   [Context Flow](https://agentic-atlas.dev/nodes/context-flow)
   how context enters, stays in, and leaves a model's window, and how well it's recalled
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/momentum#relationships) · occurrence 1
4. in-slice · occurrence 2
   [LLM Dispatch](https://agentic-atlas.dev/nodes/llm-dispatch)
   one LLM execution starting another, whether the parent waits, joins later, or never joins
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/momentum#relationships) · occurrence 2
5. undisclosed · occurrence 3
   Undisclosed relationship
   validate at the return seam; re-dispatch over repair; bounded autonomy — the contract keystone, applied
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/momentum#relationships) · occurrence 3
6. in-slice · occurrence 4
   [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment)
   how well the available context fits the task at hand
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/momentum#relationships) · occurrence 4
7. in-slice · occurrence 5
   [The Context Economy](https://agentic-atlas.dev/nodes/context-economy)
   what each piece of context costs, in quality, dollars, and latency, and who pays
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/momentum#relationships) · occurrence 5
8. undisclosed · occurrence 1
   Undisclosed relationship
   state the handle and purpose immediately before a governed act — momentum's good pole deliberately seeded, plus a legible receipt
   Evidence: [Open questions / TODO](https://agentic-atlas.dev/nodes/momentum#open-questions-todo) · occurrence 1
9. in-slice · occurrence 2
   [LLM Dispatch](https://agentic-atlas.dev/nodes/llm-dispatch)
   one LLM execution starting another, whether the parent waits, joins later, or never joins
   Evidence: [Open questions / TODO](https://agentic-atlas.dev/nodes/momentum#open-questions-todo) · occurrence 2
10. undisclosed · occurrence 3
    Undisclosed relationship
    validate at the return seam; re-dispatch over repair; bounded autonomy — the contract keystone, applied
    Evidence: [Open questions / TODO](https://agentic-atlas.dev/nodes/momentum#open-questions-todo) · occurrence 3

### Inbound references 9

1. in-slice · occurrence 4
   [Context Alignment](https://agentic-atlas.dev/nodes/context-alignment#relationships)
   how context already in the window steers the model along its current path
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/context-alignment#relationships) · occurrence 4
2. in-slice · occurrence 4
   [Context Efficiency](https://agentic-atlas.dev/nodes/context-efficiency#scope-and-boundaries)
   how context already in the window steers the model along its current path
   Evidence: [Scope and boundaries](https://agentic-atlas.dev/nodes/context-efficiency#scope-and-boundaries) · occurrence 4
3. in-slice · occurrence 1
   [Context Flow](https://agentic-atlas.dev/nodes/context-flow#model-and-claims)
   how context already in the window steers the model along its current path
   Evidence: [Model and claims](https://agentic-atlas.dev/nodes/context-flow#model-and-claims) · occurrence 1
4. in-slice · occurrence 6
   [Context Flow](https://agentic-atlas.dev/nodes/context-flow#relationships)
   how context already in the window steers the model along its current path
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/context-flow#relationships) · occurrence 6
5. in-slice · occurrence 1
   [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window#context)
   how context already in the window steers the model along its current path
   Evidence: [Context](https://agentic-atlas.dev/nodes/same-shards-fresh-window#context) · occurrence 1
6. in-slice · occurrence 1
   [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window#choice)
   how context already in the window steers the model along its current path
   Evidence: [Choice](https://agentic-atlas.dev/nodes/same-shards-fresh-window#choice) · occurrence 1
7. in-slice · occurrence 1
   [Same Shards, Fresh Window](https://agentic-atlas.dev/nodes/same-shards-fresh-window#verification)
   how context already in the window steers the model along its current path
   Evidence: [Verification](https://agentic-atlas.dev/nodes/same-shards-fresh-window#verification) · occurrence 1
8. in-slice · occurrence 5
   [Statelessness](https://agentic-atlas.dev/nodes/statelessness#implications-the-consequence-map)
   how context already in the window steers the model along its current path
   Evidence: [Implications](https://agentic-atlas.dev/nodes/statelessness#implications-the-consequence-map) · occurrence 5
9. in-slice · occurrence 4
   [Statelessness](https://agentic-atlas.dev/nodes/statelessness#relationships)
   how context already in the window steers the model along its current path
   Evidence: [Relationships](https://agentic-atlas.dev/nodes/statelessness#relationships) · occurrence 4

[↑ back to the top](https://agentic-atlas.dev/nodes/momentum#content) [← the survey](https://agentic-atlas.dev/atlas)

Node momentum · corpus 78c0e17 · Catalog revision e0cb75881244b1a82193ca738b82a0d508dd62e822524ae82873ec79d51dbb61