{"id":2955,"date":"2026-09-07T03:10:14","date_gmt":"2026-09-07T03:10:14","guid":{"rendered":"https:\/\/dmarketertayeeb.com\/blog\/gpt-6-astra-context-management-notes\/"},"modified":"2026-09-10T03:57:07","modified_gmt":"2026-09-10T03:57:07","slug":"gpt-6-astra-context-management-notes","status":"publish","type":"post","link":"https:\/\/dmarketertayeeb.com\/blog\/gpt-6-astra-context-management-notes\/","title":{"rendered":"GPT-6 Astra Context Management: How to Use Notes Without Trusting Stale Memory"},"content":{"rendered":"\n

GPT-6 Astra’s experimental Codex context management lets eligible users keep notes across context windows and search earlier messages and tool results from the same task.<\/strong> It is off by default, requires a new task, and does not guarantee that every note is current or that a long session will use fewer tokens. Treat it as a feature to verify with a disposable facts fixture, not as a promise of perfect memory.<\/p>\n\n\n\n

This guide explains the current setup, the distinction between standard compaction and experimental context management, and a practical test for exact retrieval and stale-note correction. It complements DMT’s broader GPT-6 Astra prompting guide<\/a> and the older AI-agent harness context-compaction article<\/a>.<\/p>\n\n\n\n

What the experimental feature does<\/h2>\n\n\n\n

OpenAI’s current Codex model documentation<\/a> says that supported Codex clients let ChatGPT Plus and Pro users opt in. Astra can keep notes across context windows and search earlier messages and tool results from the same task. The experiment is off by default and, at launch, is unavailable with Business, Enterprise or API-key sign-in.<\/p>\n\n\n\n

The open-source Codex implementation change<\/a> describes the setting as under development. It enables token-budget context, history notes and a model-controlled new_context<\/code> tool for eligible sessions. That implementation detail explains the design direction; it is not a guarantee that the model will write a useful note, retrieve it at the right time or resolve a conflict with a changed file.<\/p>\n\n\n\n

Question<\/th>Documented answer<\/th>What remains to test<\/th><\/tr><\/thead>
Is it enabled automatically?<\/td>No. The experiment is off by default.<\/td>Whether your current client and account actually activate it.<\/td><\/tr>
How do I enable it?<\/td>Set features.context_management.experimental_mode = true<\/code> in config.toml<\/code>, then start a new task.<\/td>Whether the new task reports the expected context tools\/behavior.<\/td><\/tr>
What can it search?<\/td>Earlier messages and tool results from the same task.<\/td>Whether the model retrieves the exact source needed at the point of use.<\/td><\/tr>
Does it make stale notes impossible?<\/td>No such guarantee is documented.<\/td>Whether a changed source file supersedes an older note.<\/td><\/tr>
Does it guarantee lower cost?<\/td>No.<\/td>Actual usage, latency, retries and recovery work on your task.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Enable it without mistaking visibility for activation<\/h2>\n\n\n\n

Configuration correction, 10 September 2026:<\/strong> use the nested table below, or merge the relative dotted key into an existing [features]<\/code> table. A root dotted assignment followed later by [features]<\/code> can redeclare that table and fail parsing. Putting the dotted line before the first table is therefore not sufficient advice for every existing configuration.<\/p>\n\n\n\n

[features.context_management]\nexperimental_mode = true<\/code><\/pre>\n\n\n\n
If [features.context_management]<\/code> already exists, edit its current value. Alternatively, inside an existing [features]<\/code> table, add context_management.experimental_mode = true<\/code>. Choose one form; do not create duplicate tables or place the key inside an unrelated table.<\/p>\n\n\n\n
  1. Confirm a supported client and eligible ChatGPT sign-in. The Models page lists Plus\/Pro; the current configuration reference also names Pro Lite. Verify your actual client\/account while that wording differs.<\/li>
  2. Back up and edit the configuration used by that client. The usual file is ~\/.codex\/config.toml<\/code>; check a custom CODEX_HOME<\/code> and more-specific project or CLI overrides.<\/li>
  3. Validate the file and confirm that the actual client recognizes the feature. codex --version<\/code> identifies the CLI; codex features list<\/code> reports its feature state.<\/li>
  4. Start a new task, as the official documentation requires. Record client, date, model, effort and config state.<\/li>
  5. Use a disposable facts fixture to test a real transition and a changed source before assuming useful retention or lower usage.<\/li><\/ol>\n\n\n\n
    In the September 10 checks, the desktop runtime 0.153.4<\/code> accepted the feature while an older PATH CLI 0.140.0<\/code> rejected its nested map. Align the command with a supported runtime, or put the experiment in a separate supported-client profile and remove it from the shared base config. An override of false<\/code> will not cure an old client’s inability to parse the table. Current profile files are separate profile-name.config.toml<\/code> files, as described in OpenAI’s profile documentation<\/a>.<\/p>\n\n\n\n
    Four local fixtures were executed with Python’s tomllib<\/code>: a root dotted flag followed by [features]<\/code> failed; a relative key inside [features]<\/code> and a dedicated nested table resolved correctly; a misplaced key under another table could parse at the wrong scope. These checks establish syntax and placement only. See the usage case study and setup-check screenshot<\/a> for the measured context traces and the client-recognition distinction.<\/p>\n\n\n\n
    Seeing Astra in a picker does not prove that experimental context management is active. The Models documentation<\/a> excludes Business, Enterprise and API-key sign-in at launch. Recheck current feature eligibility<\/a> and configuration precedence<\/a> as rollout changes.<\/p>\n\n\n\n
    A safe facts fixture for a real transition test<\/h2>\n\n\n\n
    A useful test needs a fact that is easy to verify, a source that can change, and a clear rule for which source wins. The following fixture is an original proposal for a local disposable folder. It has not been run by DMT.<\/p>\n\n\n\n
    # facts.md \u2014 proposed test fixture\nRelease channel: staging\nApproval owner: reviewer@example.test\nRule version: 1\nSource of truth: this file\nLast changed: 2026-09-07<\/pre>\n\n\n\n
    Use a task that reads the file, records the exact values and performs a harmless local action such as creating a report. Continue long enough that a genuine context-window transition occurs. Record the transition rather than assuming it happened because the task became long. After the transition, ask for the exact release channel, approval owner, rule version and source path. The grader should compare the answer with the file, character for character where appropriate.<\/p>\n\n\n\n
    Then edit only the named source fields\u2014Release channel<\/code>, Rule version<\/code> and Last changed<\/code>. Preserve Approval owner<\/code> and Source of truth<\/code> so the test changes one fact and one version rather than silently replacing the whole fixture:<\/p>\n\n\n\n
    Release channel: production\nRule version: 2\nLast changed: 2026-09-07<\/pre>\n\n\n\n
    Ask the agent to report the current values and explain what happened to the old note. A passing freshness behavior would use the current file, identify the old version as superseded and avoid silently repeating \u201cstaging.\u201d An old note is not itself a failure if the agent re-reads the source and explicitly supersedes it. The failure is using the old value as current after the source changed.<\/p>\n\n\n\n
    The result sheet<\/h2>\n\n\n\n
    Use one row per task run. Do not report a percentage until the denominator and the failed examples are preserved.<\/p>\n\n\n\n
    Field<\/th>Record<\/th><\/tr><\/thead>
    Run identity<\/td>Run ID, task fixture version, client, date and fresh\/new-task status.<\/td><\/tr>
    Model conditions<\/td>Exact model, reasoning setting, tools, experimental flag and plan\/client surface.<\/td><\/tr>
    Transition evidence<\/td>Whether a context-window transition or standard compaction was actually observed, with available log evidence.<\/td><\/tr>
    Exact recovery<\/td>Each requested fact, expected value, returned value and pass\/fail.<\/td><\/tr>
    Supersession<\/td>Whether the changed source won over the older note, and whether the model explained the change.<\/td><\/tr>
    Operational cost<\/td>Exposed usage, elapsed time, tool calls, retries and human correction time; mark unavailable fields as unavailable.<\/td><\/tr>
    Failure class<\/td>Missing retrieval, stale source, wrong task, compaction hang, tool error, authority conflict or unknown.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    Run a baseline with the same fixture and task under the normal context behavior if you need a comparison. Keep model, effort, tool permissions and grader fixed. A two-by-two design can cross inherited versus cleaned instructions with default versus experimental context management, but do not change both treatments and then attribute the result to one of them.<\/p>\n\n\n\n
    What early users are reporting<\/h2>\n\n\n\n
    The current public evidence is mixed and too early for a success rate. It is useful for deciding what to test:<\/p>\n\n\n\n
    Source<\/th>Evidence class<\/th>Report<\/th>Limit<\/th><\/tr><\/thead>
    r\/codex experimental-setting thread<\/a><\/td>First-hand post<\/td>The author says the new behavior feels better, loses fewer details and repeats less after transitions; existing sessions initially had problems, while new sessions had not shown the same issue.<\/td>One author, no task count, no controlled baseline, no usage log.<\/td><\/tr>
    r\/codex starts-from-scratch report<\/a><\/td>First-hand post and comment<\/td>Users describe Astra changing direction after compaction and disregarding a staged audit workflow; one comment says an external .md<\/code> reduced the worst effects for other models.<\/td>It does not establish that experimental mode was enabled or isolate Astra from standard compaction.<\/td><\/tr>
    r\/codex auto-compact report<\/a><\/td>First-hand post<\/td>The author reports indefinite compaction across multiple chats on Astra High with default settings; comments compare slower Astra compaction with faster Sol compaction.<\/td>No controlled run or denominator; comments are separate users and may reflect a bug rather than the experiment.<\/td><\/tr>
    OpenAI Codex issue #22219<\/a><\/td>Repository issue<\/td>A reporter describes stale task-state after repeated work and says no compaction was observed during the incident.<\/td>Stale state is self-reported and has no minimal deterministic reproduction.<\/td><\/tr>
    Ether_eagle1 reply<\/a><\/td>Open question<\/td>Asks whether notes become stale after a refactor and can be superseded.<\/td>Question, not a first-hand result; it identifies the freshness test this guide adds.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    Several negative reports refer to compaction but do not show the experimental flag, while the positive report is one user’s early experience. Treat them as failure modes and test ideas. Do not call this a community consensus.<\/p>\n\n\n\n
    Design notes so stale memory can be detected<\/h2>\n\n\n\n
    The core risk is not only forgetting. An agent can remember an old fact with enough confidence to make it look current. Codex Discussion #42703<\/a> proposes source pointers and explicit update\/delete\/no-op operations so a note resolves to current evidence instead of becoming a hidden copy. The open-source coldstart project<\/a> uses content hashes and marks notes when the anchored file changes. These are design artifacts, not proof that Astra implements either approach, but they suggest concrete checks:<\/p>\n\n\n\n