TL;DR for operators

A coding agent can inspect the current implementation and still miss why an earlier developer introduced a compatibility constraint. That reasoning may survive in historical pull requests (PRs), even when it is absent from today’s source files.

Chronos1 converts merged PRs into structured experience and connects related changes through a software-evolution graph. In an expert evaluation, graph-grounded retrieval found 2.87 useful historical records per task, versus 1.24 with ordinary embedding retrieval, among ten returned records.

The resulting coding-agent workflow also improves benchmark resolution. Across six LLM backbones on SWE-Bench Verified, average resolution rises from 69.2% to 72.9%. Yet a component comparison attributes only 0.8 percentage points of the evaluated MiniMax M2.5 workflow’s 3.4-point improvement to graph retrieval itself.

For engineering operators, the practical sequence is to evaluate historical-memory retrieval first, then test whether specialized patch-generation and selection agents justify their additional expense. The study measures benchmark performance, not production cost efficiency or guaranteed fidelity of generated historical records.

A correct-looking patch can repeat an old mistake

Consider an agent repairing a regression in a mature repository. It identifies the failing function, proposes a change, and verifies that a targeted test passes. The patch may nevertheless reverse an earlier compatibility decision whose original justification no longer appears in the surrounding code.

A previous PR might explain why an unusual condition was necessary. Another might document a subsequent change in assumptions. Neither description must resemble the present bug report closely enough to appear in a conventional semantic-search shortlist.

The operational problem is therefore broader than finding relevant code: the agent must recover historical reasoning and determine whether it still applies.

Chronos investigates whether development history can supply that missing context and whether more relevant context ultimately produces better patches.

Historical changes become connected experience, not just searchable text

Chronos constructs a repository-specific memory from merged PR titles, discussions, descriptions, and patches. A language model distills each PR into an experience card with four fields:

  • Summary: the historical change, used for semantic retrieval.
  • Signals: cues that might indicate relevance to a new task.
  • Motivation: the inadequate assumption or constraint that prompted the change.
  • Solution: implementation reasoning, code patterns, and potential pitfalls.

The distinction between Motivation and Solution matters. A patch from an earlier version may no longer be appropriate, while the constraint it addressed may remain relevant.

Chronos then connects PR events through eleven relationship types. These include overlapping changed files, subsequent extensions of an implementation, explicit cross-PR references, backports, shared reviewers, and development metadata.

This graph represents relationships between historical changes, rather than static dependencies between current source-code components.

The design extends structured human-experience memory by making relationships between historical events part of retrieval. It also differs from repository-memory approaches primarily concerned with locating code: the desired output is an explanation of prior engineering decisions that can inform a new modification.

Graph retrieval finds decisions that semantic similarity misses

Retrieval begins with a semantic search over the experience-card summaries. Chronos then follows weighted graph connections from the initial matches.

Its default configuration retains up to five semantic entry cards, explores up to three graph hops with a bounded frontier, and adds up to five distinct connected candidates. Scores combine the relevance of the starting card, relationship weights, and a penalty for longer paths.

Code-evolution relationships generally receive higher weights than organizational relationships. These are hand-specified priorities, not weights learned from downstream resolution outcomes.

The retrieval evaluation provides direct evidence for the approach.

Retrieval method Useful cards per task Tasks with a useful card
Random 0.18 14%
Embedding similarity 1.24 68%
Chronos graph retrieval 2.87 89%

Source: Section 4.4, Table 7. Main retrieval-quality evaluation; 100 randomly sampled SWE-Bench Verified tasks, ten retrieved cards per method per task.

Graph retrieval supplies an additional 1.63 useful cards per task and increases the proportion of tasks receiving at least one useful card by 21 percentage points.

These are expert judgments of task usefulness, not measurements of successful code changes or independent verification that every card faithfully reproduces its source PR. The two annotators belonged to the research team, although their reported pre-adjudication agreement was high (Cohen’s kappa = 0.86).

A separate relation-removal ablation helps explain the retrieval result. Removing the precedes relation reduces average useful cards by 0.60; removing extends reduces it by 0.44; removing explicit references reduces it by 0.40. Individual organizational relations generally have smaller effects.

Those losses are conditional on the remaining graph and cannot be added together. They nevertheless suggest that continuity of code changes and explicit development references are promising starting points for a practical implementation.

Better historical retrieval produces a smaller measured patch-resolution gain

The critical distinction appears when retrieval usefulness is compared with successful task resolution.

Chronos improves resolution across all six evaluated LLM backbones on the 500-task SWE-Bench Verified benchmark. The unweighted average rises from 69.2% to 72.9%, an improvement of 3.7 percentage points.

With MiniMax M2.5, the relevant component comparisons are:

Configuration Resolution rate
Vanilla SWE-Agent 76.4%
Full workflow with flat semantic retrieval 79.0%
Full Chronos workflow with graph retrieval 79.8%

Source: Section 4.3, Table 6. Component ablation on SWE-Bench Verified with MiniMax M2.5.

Replacing flat retrieval with graph retrieval increases full-workflow resolution by 0.8 percentage points. The complete 3.4-point improvement over vanilla SWE-Agent includes other changes to how patches are generated and selected.

That is not a contradiction. Retrieving an actionable historical record and successfully translating it into a correct patch are different outcomes. The latter requires diagnosis, applicability judgments, implementation, and validation.

The broader benchmark comparisons strengthen the finding that the full workflow can help. With MiniMax M2.5, resolution increases from 48.3% to 51.7% on the public SWE-Bench Pro subset and from 41.0% to 43.5% on FEA-Bench Lite. Those two benchmarks were not evaluated across all six backbones.

Two patch strategies improve the workflow, but the selector’s value remains unisolated

Chronos adds two complementary candidate-generation agents.

The patch-focused agent investigates design intent, violated assumptions, and historical implementation patterns. The validation-focused agent emphasizes expected behavior, affected APIs, and regression risks.

When both produce distinct, non-empty patches, an independent evolution steward examines the alternatives using current code and repository history. It chooses between the existing patches; it does not produce a third.

The component comparisons report 78.4% resolution for the graph-enabled patch-focused single-agent configuration and 77.8% for the validation-focused configuration. Combining both strategies with candidate selection reaches 79.8%.

These results support the complete workflow relative to the evaluated single-agent alternatives. They do not separately establish the steward’s contribution: removing one generation strategy also removes the second candidate and the need for steward selection. Likewise, single-agent improvements combine historical retrieval with role-specific prompting.

An operator should therefore avoid treating the full-to-single-agent difference as a measured return on adding a selector.

Evaluate memory quality before paying for the entire workflow

Cognaptus inference: Engineering organizations with substantial PR histories can test the retrieval layer as an independently measurable capability before introducing multiple coding agents.

A staged evaluation would measure four distinct outcomes:

  1. Historical retrieval: Does connected PR history improve task-relevant context at a fixed retrieval depth?
  2. Source fidelity: Do generated cards accurately preserve the historical rationale, constraints, and pitfalls recorded in their source PRs?
  3. Patch quality: Does access to history improve resolution and regression outcomes against a controlled agent baseline?
  4. Incremental expenditure: Do additional generation and selection runs deliver enough improvement to justify their inference, execution, preprocessing, and maintenance costs?

Repository scoping and task-time timestamp filters are useful architectural controls. They do not, by themselves, establish a general guarantee against information leakage.

The cost boundary is also material. Chronos requires PR distillation, embeddings, graph construction, two candidate-generation runs, and potentially a steward run. The paper provides no complete end-to-end cost-efficiency or preprocessing-amortization analysis.

Five repetitions of one full MiniMax M2.5 configuration produce a mean SWE-Bench Verified resolution rate of 80.0%, with a sample standard deviation of 0.37 percentage points. This is a useful stability check, but the repeated-run interval describes Chronos alone, not the statistical uncertainty of its improvement over a repeatedly evaluated baseline.

Production reliability, historical-card fidelity, and maintenance overhead remain open deployment questions.

The first investment is recoverable engineering rationale

Chronos demonstrates that historical software changes contain retrievable information that current source code and flat similarity search may not adequately expose. Its strongest direct result is the improvement in task-useful historical retrieval; the overall coding gains additionally reflect how agents consume that history and select patches.

For an engineering organization deciding where to invest, the evidence favors establishing the value and fidelity of PR-derived memory before assuming that the full multi-agent architecture is necessary. Historical rationale can improve an agent’s investigation, but the final patch still has to satisfy the current repository’s requirements.


Cognaptus: Automate the Present, Incubate the Future.


  1. Xin Yin and Yiang Zhang and Zhiyuan Peng and Chao Ni and Zhe Cui and Xiaohua Xin (2026). Chronos Enables Code Agents to Reason over Software Evolution. arXiv:2610.11578. https://arxiv.org/abs/2610.11578 ↩︎