Skip to content

Verify: age-out synthetic Observations (upstream #1566) #471

Closed

Description

@possiblyneal

Part of #448

Question

Do importance-5 synthetic Observations older than 60 days (10,564 of 103,152 on the live store; ~24,000 older than 30 days) ever reach an Injection in the replay or decision-key probe, and would evicting them lose a key item? Upstream: rohitg00#1566. Measure only; a change to default retention goes to the Operator with the numbers (Durable beats recent vs disk growth). Evidence: tmp/verify/470/triage.md.

Activity

  1. possiblyneal commented on Nov 18, 2026

    @possiblyneal
    OwnerAuthor

    Live-store reach, read-only copy (tmp/verify/live-copy.sqlite, scripts tmp/verify/471/reach.cjs, access2.cjs). Replay/decision-key probe still to run.

    Delivered (mem:injections, 2026-11-11 → 11-18, 3,533 Injections, 2,967 Observation refs):

    surface compressed <7 d 7–30 d 30–60 d synthetic <7 d synthetic 7–30 d synthetic ≥30 d
    session-start 1,516 0 0 20 0 0
    prompt-submit 344 26 1 55 3 0
    pre-compact 612 0 0 7 0 0

    No synthetic Observation 30 days or older was injected in the logged week (383 refs point at Observations already evicted).

    Read as a search hit (mem:access, since 2026-05-20; counts candidates, not only deliveries): 1,111 of 10,125 synthetic Observations ≥60 days old were hit at least once after day 60 (3,447 hits); 537 of them inside the logged week, in which none was injected. So late hits are search candidates the floor, gate or rank dropped, or agent memory_smart_search results; the access log does not say which.

    Next: decision-key probe with and without the ≥60-day synthetic set, to see whether any key item depends on one.

  2. possiblyneal commented on Nov 18, 2026

    @possiblyneal
    OwnerAuthor

    Verdict: not applied (measured, no Operator decision needed)

    Replay cannot test it. The frozen decision-key corpus (tmp/eval-arms/facts-a6f2bd, 29 Sessions, 421 decisions) spans 2026-09-26 → 10-26, 30 days; no Observation in it reaches 60 days, so the probe cannot show a key item lost to a 60-day rule. Claude Code keeps 30 days of transcripts, so a fresh corpus would not either.

    Disk the rule would free (read-only copy, tmp/verify/471/bytes.cjs): store 598 MB, Observations 185 MB JSON + 154 MB vectors.

    rule rows JSON vectors share of store
    synthetic ≥60 d 10,125 13.1 MB 14.7 MB ~4.6%
    synthetic ≥30 d 22,922 37.1 MB 31.7 MB ~11.5%

    What it would remove (previous comment): no old synthetic Observation was injected in the logged week, but 1,111 of the ≥60-day set were read as search hits after day 60 (3,447 hits since May), and the log cannot separate gate-dropped candidates from Agent memory_smart_search results.

    Under the Goal, Volume is a cost but a ~28 MB gain does not outweigh deleting records some of which are still read; Session Summaries already carry the covered ones. No change to default retention, so nothing goes to the Operator.

    Retry if: the store passes ~2 GB or disk growth shows up in /health, or the access log gains a surface field that shows the late hits are gate-dropped candidates only.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions