Skip to content

Adopt competitive program: PR1 LME-20 pin + PR2 write policy - #100

Merged
cursor[bot] merged 15 commits into
devfrom
pr/competitive-program-adoption-a6c7
Aug 13, 2026
Merged

Adopt competitive program: PR1 LME-20 pin + PR2 write policy#100
cursor[bot] merged 15 commits into
devfrom
pr/competitive-program-adoption-a6c7

Conversation

@sdntsng

@sdntsng sdntsng commented Aug 11, 2026 •
edited by cursor Bot
Loading

Copy link
Copy Markdown
Contributor

Summary

Adopts the competitive architecture review as the program of record (ACCEPT with modifications), then executes PR1 (LME-20 measurement integrity) and PR2 (conversational vs governed write policy).

Strategic frame: recall broadly → represent explicitly → prove narrowly → answer truthfully.

Merge target is dev only. Do not merge or push to main / production unless the owner explicitly asks.

Honest pins only. No SOTA / beats-Mem0 claims. LME-20 0/20 is an integrity pin, not a quality win. LoCoMo remains post-cutover 15/30 and 33/90.

Changes

Phase 0 — program adoption (docs)

  • Verdict, competitive SOP / gap matrix, handoff surfaces

PR1 — LME measurement integrity (done)

  • Harness: failure_reason + job accounting; proveability fail-closed
  • UTF-8 sanitize (ingest + store write-boundary)
  • Dedupe idempotent async job_ids (attempt-2 off-by-one 4829 != 4830)
  • Write run json/manifest/report before proveability SystemExit
  • Publishable pin lme20-product-recall-pr1-20260812:
    • --publish --product-recall --limit 20 --seed 1 --async-timeout 3600
    • Proveability passed: answer_path=/recall 20/20; jobs_expected=jobs_completed=4829; jobs_failed=0
    • Accuracy 0/20 (0.000) — honest, not a quality claim
    • Dataset SHA d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442
    • Pin: docs/benchmarks/artifacts/lme20-product-recall-pr1-20260812-pin.md
    • Run manifests stay gitignored (secret scan)

PR2 — conversational append-only vs governed mutation (code landed)

  • WriteMutationModeOf: core/empty vertical → append-only; any non-core vertical → governed (fix: provider NONE/DELETE/UPDATE authoritative over baseline #94 ops kept)
  • Conversational merge no longer drops baseline on provider NONE/UPDATE/DELETE (NY history retained after “moved to Austin”)
  • Conversational UPDATE/DELETE persist as ADD; no extract-time suppress/supersede
  • Support/marketing (and other non-core verticals) still suppress/supersede
  • Explicit /correct, /suppress, /supersede APIs unchanged
  • Local non-reg after rebuild: OpMem 13/13, marketing passed
  • No LoCoMo / LME quality remasure yet — do not invent lift

Test plan

  • UTF-8 Go tests
  • Job-id dedupe harness test
  • Clean publishable LME-20 pin (0/20 integrity)
  • go test ./internal/memory/ ./internal/jobs/ ./internal/api/
  • Conversational UPDATE retains prior; governed UPDATE supersedes
  • Conversational DELETE does not suppress; governed DELETE does
  • Local OpMem 13/13 (pin)
  • Local marketing passed (pin)
  • LoCoMo 1×30 / 3×90 remasure after PR2 (next; honest pin only)
  • LME-20 quality remasure after conversational lift (not a claim now)

Linear / tracking

Competitive program of record: docs/research/external-reviews/2026-08-11-competitive-architecture-verdict.md
Status: docs/research/program-execution-status.md

Risk / follow-ups

  • LME-100 measurement gate is open; quality is not ready (0/20)
  • Auto-supersede of stateful predicates on conversational ingest is unchanged (PR3 temporal ranking still queued)
  • Next: PR3 temporal features V1 (temporal_score, reuse mig-16), then PR4–PR10
  • Same-pin Mem0 qualification only after LoCoMo + publishable LME-20 quality lift
  • This PR is large; prefer a follow-up branch to dev for PR3+
  • Never main without an explicit owner ask
Open in Web Open in Cursor 

sdntsng and others added 8 commits August 10, 2026 11:59
@sdntsng
Merge: external re-review brief onto main (production)
6b307a4 
Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
Merge: V3 hardening onto main (production)
308d3a1 
Land subject-ordered claims, authoritative ops, LME product-recall
harness, hop executor V2, truthful hybrid status, and Gate 0 /
qualification docs from staging (dev@1f2f26f).

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
docs: revise external review guidelines + hardening self-review prompt
7cda664 
Point intake SOP and assessment pack at the post-cutover state (hardening
merged to dev/main), refresh claims discipline, and add a dedicated
copy/paste self-review prompt for the external reviewer. Record post-cutover
OpMem/marketing staging pins and LME abort honesty.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
docs: pin post-cutover staging LoCoMo 1x30 (15/30)
a6b50f4 
Record staging deploy 1f2f26f LoCoMo 1x30 at 15/30 with honest compare
vs Gate 0 18/30 and harden-local 14/30; refresh handoff surfaces.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
docs: pin post-cutover staging LoCoMo 3x90 (33/90)
aba013c 
Complete full-pass pins on deploy 1f2f26f and refresh external-review
handoff surfaces with honest Gate 0 vs post-cutover compares.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
docs: adopt competitive architecture program as PoR
2f2ae51 
Archive ACCEPT-with-modifications verdict, add competitive archaeology
SOP/gap matrix/borrow log, and point handoff surfaces at PR1-PR10.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
fix: LME publish job accounting + UTF-8 sanitize
9e2925b 
Surface extraction failure_reason in publish-mode job waits, require
jobs_completed==jobs_expected with jobs_failed==0, and strip invalid
UTF-8 before persist (root cause of LME-20 abort SQLSTATE 22021).

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
fix: sanitize UTF-8 at store write boundary
4e56229 
Defense-in-depth for LME SQLSTATE 22021: sanitize text fields in
UpsertMemory, atoms/events/entity hub, and evidence writers so provider
extract paths cannot insert invalid byte sequences even if an earlier
normalize step is skipped.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@cursor cursor Bot changed the title docs+harness: adopt competitive architecture program (PR1 LME integrity) Adopt competitive architecture program + LME-20 measurement integrity (PR1) Aug 11, 2026
sdntsng and others added 6 commits August 11, 2026 13:30
@sdntsng
docs: pin LME-20 PR1 attempt-1 timeout (UTF-8 gate cleared)
eaa6e7e 
Record fail-closed abort on item 17 async barrier timeout after 16/20
clean items; note prior UTF-8 abort subject 4388e9dd now completes.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
fix: dedupe LME async job_ids + pin PR1 attempt-2 artifact
042da72 
Idempotent /ingest/async could return a duplicate job_id that inflated
jobs_expected by 1 and fail-closed a full 20/20 /recall LME-20 run.
Deduplicate pending IDs in the harness; pin the attempt-2 failure
artifact (UTF-8 and timeout gates cleared; proveability blocked).

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
fix: write LME run artifacts before proveability fail-closed
1225be5 
Attempt-2 scored 20/20 /recall but exited before json/manifest/report
were written. Pin evidence first, then SystemExit on gaps.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
docs: pin publishable LME-20 product-recall 0/20 (PR1)
d1ac91d 
Proveability passed: 20/20 /recall, jobs 4829=4829 failed=0.
Accuracy 0/20 is honest integrity, not a quality win.
Do not commit run manifests (env URLs trip secret scan).

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
feat: conversational append-only vs governed write policy (PR2)
d5d49cc 
Core/empty vertical keeps provider NONE/UPDATE/DELETE from dropping
or mutating prior conversational records. Non-core verticals retain
#94 authoritative ops. No LoCoMo/LME quality claim.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
test: governed write-policy fixtures use conversation source_type
10a31e3 
crm source_type skips deterministic episode extract, so seed ingest
created zero memories. Vertical=support still selects governed mode.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@cursor cursor Bot changed the title Adopt competitive architecture program + LME-20 measurement integrity (PR1) Adopt competitive program: PR1 LME-20 pin + PR2 write policy Aug 13, 2026
@sdntsng
docs: pin PR2 local OpMem 13/13 and marketing non-reg
d140163 
Rebuilt local API from this branch; no LoCoMo/LME quality claim.

Co-authored-by: Siddhant Singh <sdntsng@users.noreply.github.com>
@sdntsng
sdntsng marked this pull request as ready for review August 13, 2026 11:19
@cursor
cursor Bot merged commit 24be5ab into dev Aug 13, 2026
2 checks passed
This was referenced Aug 13, 2026
Merged
Merged
Merged
Merged
Merged
@cursor cursor Bot mentioned this pull request Aug 13, 2026
Merged
6 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@sdntsng