Watch & compare
Memory architecture
Different ways to see the same idea. Play one cut at a time, then return to the guide to try it yourself.
Why we recommend Recall is not current
Start with Astra: it makes a recalled command visibly succeed at retrieval, fail a source-revision check, and become usable only after both saved memory and session context are refreshed. Baseline complements it by showing an already active context staying stale after storage changes, while keeping another thread’s notes separate. Fable adds the write-side lesson: one unreviewed guess is reused across runs until an application gate and an approved replacement change later behaviour; its denser seven-run story is better as a follow-up.
Revised original · 26 sec
Scope and explicit refresh
Keep thread notes separate from shared conventions, and explicitly retrieve a corrected record into an active context.
Read the video sequence
- 0–9s: Thread-local notes remain separate from a durable convention stored with an approved source revision. Agent A retrieves that convention.
- 10–15s: Agent B starts with empty context and retrieves the durable convention rather than the other thread’s temporary notes.
- 16–22s: A source correction updates storage. B still has the old version until it retrieves the corrected value.
- 22–26s: Hold the refreshed active context at r2, then fade the complete image to black.
About this cut
Production: Original prototype; model not recorded.
Revision: gpt-6-astra comparative review.
Scope is separate from durability: Thread A can remain checkpointed while B is active.
Amber highlights version divergence for the viewer; this cut does not implement a freshness validator or permission check.
The approved source and correction are illustrative; no repository integration was executed.
Audio is measured analytically, not listened to.
Recommended · Astra · 28 sec
Recall is not current
Check a remembered repository instruction against its source revision, then refresh both the stored record and the current context before acting.
Read the video sequence
- An approved repository test command at revision r1 is saved, then recalled into session A. Both use npm test.
- The repository changes to revision r2 and pnpm test. Saved memory still holds r1. A new session B opens without a command.
- Session B recalls the old command successfully. A separately implemented source revision check detects r1 versus r2 and stops the candidate before execution.
- The workflow clears the stale record and task candidate, reloads r2 from the repository into saved memory, and supplies r2 to session B.
- Session B runs the current command and shows an illustrative passing result, then the picture fades. Correct memory does not guarantee passing tests or enforce permissions.
About this cut
Production: gpt-6-astra.
Revision checking and invalidation are illustrative application logic, not automatic store guarantees.
The passing test result is illustrative; no agent integration or real test migration was executed.
Persistent instructions do not enforce tool permissions.
Audio was analysed for levels and cue alignment, not physically listened to.
Self-reviewed; comparative selection and human editorial promotion remain separate.
Alternative · Fable · 32 sec
Gate what gets written
An unreviewed guess can repeatedly steer later runs to skip tests. An application write gate rejects that guess, keeps it in the current thread, and accepts a correction from an approved source for future retrieval.
Read the video sequence
- 0–5.6s: Run 1 puts retry x3 in Thread 1, reads lint: strict @r12 from durable memory and shows an illustrative OK. The display then switches to Run 2 and a new empty Thread 2; this does not imply that saved checkpoints were deleted.
- 5.6–10.2s: Run 2 writes flaky: ci to its thread, then writes the unsourced tests: skip ? directly into durable memory. The record remains when Run 3 starts.
- 10.2–19.8s: Runs 3, 4 and 5 retrieve skip ? and show NO TESTS. Three red dots mark these illustrative skipped-test outcomes, not a claim of deterministic agent behaviour or failing test execution.
- 20.3–23.6s: A write gate and approved TESTS.md r13 source appear. The agent’s unsourced guess stops above the gate, bounces back and is kept in Thread 6.
- 24.1–27s: A green tests: run @r13 record travels from the approved source to the gate, passes, and replaces the old memory row. SAVED marks the completed write; it does not report an executed test.
- 27–32.2s: Run 7 and Thread 7 appear. The new run retrieves run @r13, then shows RUN TESTS. The result is held before the complete image fades to black.
About this cut
Production: anthropic/claude-fable-5-1.
Revision: gpt-6-astra comparative review.
Illustrative application logic: source/revision metadata alone does not prove approval or enforce permissions. The displayed source is assumed to have been reviewed.
Three repeated skipped-test outcomes illustrate risk, not a model capability or guaranteed failure rate. No actual agent integration or tests were executed.
The animation switches between threads. Older thread checkpoints may persist; deletion and retention are not demonstrated.
The gate blocks future writes; the already stored bad entry remains until the explicit replacement at 25.45s.
Audio was checked by decoded peak levels, scheduled-cue window energy and spectrogram, not listened to. Comparative evidence is in reviews/comparisons/memory-architecture/.
These animations use sound effects rather than narration. The written guide explains the mechanism and its limits. Model names describe how a cut was made, not a quality ranking.
What the review found
- Astra 6.2–16.0s separates source change, successful stale recall and rejection before action. Its large three-card layout makes the guide’s test-command example easier to follow at 350px than Baseline’s four-node scope diagram or Fable’s run counter and gate.
- Baseline 10.0–14.2s keeps Thread A separate from Agent B; 17.4–21.3s shows r2 in storage while B remains amber at r1 until a separate GET returns r2. This directly teaches the active-context gap that Astra’s replacement of session A with B at 8.3s leaves outside its story.
- Fable 6.9–18.5s shows an unsourced skip-tests write surviving into three later runs. The revised NO TESTS outcomes identify the harm without implying that skipped tests ran and failed; repeated reuse is illustrative, not a guaranteed model behaviour.
- Fable 20.3–25.45s now shows approved TESTS.md r13 supplying the replacement through the gate, with packet marks clear of labels. SAVED at 25.9s and RUN TESTS at 29.0s separate write completion from the next run’s action. Source/revision metadata still requires a real approval policy.
- Astra 16.0–23.2s gives clearing, source refresh, context supply and test execution distinct steps, followed by a roughly three-second result hold. Baseline resolves at 21.3s and holds to 25.5s; Fable spends about ten seconds on repeated bad retrievals before its denser repair. That repetition teaches amplification but makes Fable less efficient as the first explanation.
- Baseline 25.5–26.0s, Astra 26.2–28.0s and Fable 31.4–32.2s all end on exact matching opening/closing pixels. The final sequences agree with their transcripts; none demonstrates authorization, retention, concurrent updates or repair of every active session, so the written guide and two-session exercise remain essential.