How Brainy is measured, and how that compares to other memory systems.
The README carries two tables: published % (industry scoreboard) and same-pin (the only fair lead/trail). This page is the full section: sourced claims, caveats, artifacts, and reproduce commands. Internal cycle write-ups live in research/competitive/cycle-closeout.md.
Honest rules
- We do not invent scores. Suites without a run are not run / no pin.
- Same-pin = same dataset SHA, same judge/answerer, same question set.
- Published vendor percents live in published-claims.md. Label metric and n. Do not treat them as same-pin lead/trail.
- 1×30 LoCoMo is measurement, not qualification. Not SOTA.
- Mem0 OSS ≠ Mem0 Platform. Graphiti OSS ≠ Zep Platform.
Industry format: one percent per public suite. Not same-pin.
| LoCoMo | LongMemEval | BEAM 1M | BEAM 10M | |
|---|---|---|---|---|
| Brainy | 11.4% full /recall (n=1540) · 70.0% 1×30 |
20% LME-20 /recall (4/20) · 4% (n=100 hist.) |
not run (40% on 100K/20q this cycle) | not run |
| Mem0 Platform | 92.5% | 94.4% | 64.1 | 48.6 |
| Zep | 75.14% | 71.2% | — | — |
| SuperMemory | 77.1% | 95% Recall@15 | — | — |
| Letta | 74.0% | — | — | — |
| Hindsight | 92.0% | 94.6% | 73.9% | 64.1% |
Sources, n, metric type, and the Zep LoCoMo dispute: published-claims.md. Peer harness: mem0ai/memory-benchmarks.
Pin date for Brainy: fresh remasure, 2026-08-15 (1b5ab3e product SHA).
Mem0 LoCoMo / OpMem / marketing freeze: this cycle. Full LoCoMo n=1540
landed at 11.4% product /recall. LME-20 landed at 4/20. BEAM 100K
landed at 8/20. LME-500 and BEAM 1M/10M not run.
| Suite | Brainy | Mem0 Platform | Graphiti OSS / Zep Platform | Stand |
|---|---|---|---|---|
| LoCoMo 1×30 overall | 21/30 (70.0%) | 11/30 (36.7%) | no same-pin | Lead this freeze; not SOTA |
| Multi-hop | 10/10 | 6/10 | no same-pin | Lead this freeze |
| Open-domain | 0/4 | 3/4 | no same-pin | Trail |
| Temporal | 11/16 | 2/16 | no same-pin | Lead this freeze |
| OpMem | 13/13 | 10/13 | no pin | Lead ops |
| Marketing vertical | 17/17 | 4/17 empirical | no pin | Lead governed vertical |
| LongMemEval-20 | 4/20 (20.0%) product /recall |
no fair pin on this harness | no pin | Lift vs 0/20 integrity; not vs published 94.4% |
| BEAM 100K | 8/20 (40.0%) search+harness | published elsewhere; not our pin | no pin | Non-reg vs hist. 8/20; 1M/10M not run |
Search p50 on the LoCoMo 1×30 pin: Brainy 175 ms local vs Mem0 Platform 492 ms. That is a harness observation, not a production SLO.
Mem0 OSS was not re-measured. Do not treat Platform 11/30 as an OSS-reproducible number.
The 2026-08-15 Mem0 1×30 freeze used v2 search, v1 add, chunk 8, no session
timestamps, and top_k 30. That is not their published LoCoMo protocol
(v3 search/add, chunk 1, unix timestamp, top_k 200). Do not write a
new lead/trail sentence from 21/30 vs 11/30 until the fair stratified 180
lands. Audit: mem0-harness-audit-2026-08-22.md.
Do not paste the published-% table into this same-pin n/N table. Those percents (and the Zep 58–84 dispute) are catalogued in published-claims.md.
The three suites in mem0ai/memory-benchmarks
are first-class here. We run our own adapters (UnifiedResult JSON); we do not
vendor their repo.
| Benchmark | What it tests | Upstream | Our runner |
|---|---|---|---|
| LOCOMO | Long multi-session dialogues; factual / multi-hop / temporal QA | snap-research/locomo · ACL 2024 | evals/public/locomo/ (--system brainy or mem0) |
| LongMemEval | Long-term extraction, temporal, multi-session | LongMemEval dataset | evals/public/longmemeval/ — this cycle 4/20 on /recall (pin); LME-500 not run |
| BEAM | Retrieval across 100K–10M token chats | HuggingFace Mohammadta/BEAM |
evals/public/beam/ — this cycle 8/20 on 100K conv-0 (pin); 1M/10M not run |
| Harness peer | Ingest → search → LLM answer/judge; UnifiedResult JSON |
mem0ai/memory-benchmarks | evals/public/schema.py · Brainy drop-in: evals/public/backends/memory_benchmarks_brainy.py |
Also outlinked (not in the comparison table): LongMemBench.
Our LoCoMo smoke defaults to one conversation × 30 questions, categories
1–4 (adversarial excluded), judge temperature 0. Dataset SHA for the frozen
compare: 79fa87e90f04081343b8c8debecb80a9a6842b76a7aa537dc9fdf651ea698ff4.
Reproducible in this repo against a live Brainy API (stdlib Python, no pip).
| Suite | Spec / fixtures | Runner |
|---|---|---|
| OpMem | opmem-spec · fixtures/opmem/ |
evals/run_opmem.py |
| Parity | fixtures/parity/ |
evals/run_eval.py |
| Marketing vertical | fixtures/vertical/marketing/ |
evals/run_vertical_eval.py |
| Marketing MVP / moat | METHODOLOGY | evals/run_marketing_mvp_benchmark.py |
CI (go test ./...) runs the HTTP harnesses. Same-pin vs Mem0 Platform needs
MEM0_API_KEY and is not a merge gate.
Start the API (docker compose up --build or go run ./cmd/api), then:
export BRAINY_BASE_URL=http://localhost:8080
python3 evals/run_opmem.py --systems brainy,verbatim --base-url "$BRAINY_BASE_URL"
python3 evals/run_vertical_eval.py --base-url "$BRAINY_BASE_URL"
python3 evals/run_marketing_mvp_benchmark.py --base-url "$BRAINY_BASE_URL"
# Mem0 Platform counter-run (optional)
# python3 evals/run_opmem.py --systems verbatim,brainy,mem0 --base-url "$BRAINY_BASE_URL"
# python3 evals/run_marketing_mvp_benchmark.py --base-url "$BRAINY_BASE_URL" --systems brainy,mem0
# python3 evals/run_competitor_benchmark.py --brainy-url "$BRAINY_BASE_URL"Public suites (LoCoMo 1×30 Brainy, then Mem0 Platform; LME / BEAM optional):
cd evals
python -m public.locomo.run_smoke --system brainy --conversations 1 --questions 30
# python -m public.locomo.run_smoke --system mem0 --conversations 1 --questions 30
# python -m public.longmemeval.run --limit 20 --product-recall
# python -m public.beam.run --chat-size 100K --conversations 0-0More harness detail: evals/README.md · evals/public/README.md. Ladder: research/public-bench-ladder.md.
| Suite | Report |
|---|---|
| LoCoMo 1×30 | locomo-fresh-1x30-20260815.md |
| LoCoMo full n=1540 | locomo-fresh-full-20260815.md · why 11.4% |
| LongMemEval-20 | lme20-fresh-20260815.md |
| Mem0 LoCoMo freeze | locomo-mem0-fresh-1x30-20260815.md |
| OpMem | opmem-fresh-local-20260815.md · Mem0 |
| Marketing | marketing-fresh-local-20260815.md · Mem0 · moat |
| Cycle closeout (detailed why) | cycle-closeout.md |
| Fail-closed integrity remasure (2026-08-19/20) | S0 dual-lane · 3×90 · extraction ceiling · embedding A/B. Invalidates Aug-19 17/180 / 52/180. Not a replacement for the 1×30 freeze above. |
| Current-SHA S0 this-VM (2026-08-22) | Tenant diag-mh-135 + conv-30, hybrid off: product 19/180, industry 62/180. pin. Not a replacement for integrity 32/180 or the 1×30 freeze. |
| P1 hybrid-on S0 this-VM (2026-08-22) | Same store, BRAINY_RECALL_LLM=1: product 37/180. MH 10/33 dip vs reader-off 12/33. SH 19/98, temporal 7/38. pin. Not n=1540, not a Mem0 same-pin. |
| P2-narrow S0 this-VM (2026-08-22) | Same store, hybrid on. Length-lock 56/180 (MH 11/33, temporal 21/38). pin. Extras-lock + skip hop-ground 61/180 (MH 16/33, temporal 19/38 dip vs 21/38). pin. Not n=1540, not a Mem0 same-pin. |
| P3 distinctive-token S0 this-VM (2026-08-22) | Same store, hybrid on. Admit leftover query tokens; skip unproven hop dumps. 73/180 (MH 16/33 held, OD 3/11, SH 32/98, temporal 22/38). pin. Not n=1540, not a Mem0 same-pin. |
| P4 identity/garbage S0 this-VM (2026-08-22) | Same store, hybrid on. Reject garbage hybrid; skip identity-only hop dumps; do not hop-ground those dumps. 79/180 (MH 16/33 held, OD 3/11, SH 37/98, temporal 23/38). pin. Not n=1540, not a Mem0 same-pin. |
| P5 activity-dump S0 this-VM (2026-08-22) | Same store, hybrid on. Skip activity/event hop dumps that miss leftover query tokens. 84/180 (MH 17/33, OD 2/11 dip, SH 45/98, temporal 20/38 dip). pin. Not n=1540, not a Mem0 same-pin. |
| P6 dump-lock S0 this-VM (2026-08-22) | Same store, hybrid on. Skip dual-entity activity dumps; unlock hop-dump locks; place-fact ranking. 87/180 (MH 13/33 dip, OD 3/11, SH 52/98, temporal 19/38 dip). pin. Not n=1540, not a Mem0 same-pin. |
| P7 hop-local joins S0 this-VM (2026-08-22) | Same store, hybrid on. Leftover hop contents under skip; rare-share omitted possessions; keep typed joins on hybrid abstain. 88/180 (MH 14/33, OD 4/11, SH 49/98 dip, temporal 21/38). pin. Not n=1540, not a Mem0 same-pin. |
| P8 SH recovery S0 this-VM (2026-08-22) | Same store, hybrid on. Where-only mh_list unlock; dated ordinal names; attended-event hop drop; leftover specific facts with conflicting date tails stripped. 93/180 (MH 15/33, OD 4/11, SH 52/98, temporal 22/38). pin. Not n=1540, not a Mem0 same-pin. |
| P9 unproven mh_list dumps S0 this-VM (2026-08-22) | Same store, hybrid on. Unlock unproven search_fallback mh_list dumps; skip OCR captions and question-echo hop values. 94/180 (MH 15/33, OD 4/11, SH 53/98, temporal 22/38). pin. Not n=1540, not a Mem0 same-pin. |
| P10 date-aware leftover covering S0 this-VM (2026-08-23) | Same store, hybrid on. Far-dated leftover covering skip (10-day window); hybrid 48h window except where-queries; speaker-prefix leftover covering uses the body. 96/180 (MH 15/33, OD 4/11, SH 55/98, temporal 22/38). pin. Not n=1540, not a Mem0 same-pin. |
| P11 locative leftover covering S0 this-VM (2026-08-23) | Same store, hybrid on. Where leftover covering ignores hop slots; strong leftover tokens; hybrid leftover overwrite only on where/games-played joins. 97/180 (MH 13/33 dip, OD 3/11 dip, SH 58/98, temporal 23/38). pin. Not n=1540, not a Mem0 same-pin. |
| P12 keep short where NPs / typed item joins S0 this-VM (2026-08-23) | Same store, hybrid on. Where covering returns the locative place NP; short typed item joins stay locked; leftover support is a weak token. 101/180 (MH 15/33, OD 4/11, SH 58/98, temporal 24/38). pin. Not n=1540, not a Mem0 same-pin. |
| P13 gated leftover-vs-hybrid S0 this-VM (2026-08-23) | Same store, hybrid on. Leftover covering may replace hybrid only for schema-activity covering (enjoys/likes/loves/participates in) or where/games-played; chat turns do not get a locative leftover bonus. 102/180 (MH 15/33, OD 4/11, SH 59/98, temporal 24/38). pin. Not n=1540, not a Mem0 same-pin. |
| P14 childhood possession lock S0 this-VM (2026-08-23) | Same store, hybrid on. 2-item childhood possession lists lock against speaker chat; leftover covering can join had a + age-cue facts; name questions are not childhood item lists. 103/180 (MH 16/33, OD 4/11, SH 59/98, temporal 24/38). pin. Broad typed-join 180 98/180 is not a pin. Not n=1540, not a Mem0 same-pin. |
| P17 when-event leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. When-event leftover covering keeps 4-character event verbs; a short calendar date that misses those tokens yields to a packet covering line that has them. 104/180 (MH 16/33, OD 4/11, SH 60/98, temporal 24/38). pin. P15/P16 103/180 net-zero 180s are not pins. Named dip: Jon banker job. Not n=1540, not a Mem0 same-pin. |
| P18 when-event query-entity bind S0 this-VM (2026-08-25) | Same store, hybrid on. When-event leftover covering skips lines that name another person and do not name a query person; first-person dated lines still compete. 105/180 (MH 16/33, OD 4/11, SH 60/98, temporal 25/38). pin. Jon banker recovered. Not n=1540, not a Mem0 same-pin. |
| P20 enumerate unwind extras S0 this-VM (2026-08-25) | Same store, hybrid on. Unwind-evidenced packet/hop activity slots join onto enumerate items (not only answer). 106/180 (MH 17/33, OD 4/11, SH 60/98, temporal 25/38). pin. Destress pottery recovered. P19/P19b 105/180 holds are not pins. Not n=1540, not a Mem0 same-pin. |
| P21 location-list leftover lock S0 this-VM (2026-08-25) | Same store, hybrid on. Location-list leftover covering requires a locative; leftover packet practice places join onto answer/items; which/what-year covering binds to the query person and needs a year token. 107/180 (MH 18/33, OD 4/11, SH 60/98, temporal 25/38). pin. Yoga practice locations recovered. First P21 180 106/180 flake is not a pin. Not n=1540, not a Mem0 same-pin. |
| P22 leftover month bind S0 this-VM (2026-08-25) | Same store, hybrid on. When leftover covering tokens collapse to weak-only, bind covering to the query month/year; rarest-token override ignores weak tokens. 108/180 (MH 18/33, OD 4/11, SH 60/98, temporal 26/38). pin. September co-participant plan recovered. Not n=1540, not a Mem0 same-pin. |
| P23 year/month leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. When-event leftover covering requires a year/date token so year-only event facts can replace a bare hop date; covering from a different query year or month is skipped. 109/180 (MH 18/33, OD 4/11, SH 60/98, temporal 27/38). pin. Hometown community-center 2022 recovered. First P23 180 108/180 hold is not a pin. Not n=1540, not a Mem0 same-pin. |
| P24 sentence-initial verb covering S0 this-VM (2026-08-25) | Same store, hybrid on. Sentence-initial past-tense verbs and phrasal particles are not leftover covering people, so unnamed dated diary lines still compete. 110/180 (MH 18/33, OD 4/11, SH 60/98, temporal 28/38). pin. Unnamed dated art-show diary recovered. Not n=1540, not a Mem0 same-pin. |
| P25 which-year as-of duration S0 this-VM (2026-08-25) | Same store, hybrid on. Which-year leftover covering of the form for N years as of DATE rewrites to the start year; leftover tokens year/years are weak. 111/180 (MH 18/33, OD 4/11, SH 60/98, temporal 29/38). pin. Health start year recovered. Not n=1540, not a Mem0 same-pin. |
| P26 when-event speech-act flood S0 this-VM (2026-08-25) | Same store, hybrid on. When/which-year lexical search drops leftover-weak tokens (decide/year) when another event noun remains; leftover covering prefers lines that name both query people. 112/180 (MH 18/33, OD 4/11, SH 60/98, temporal 30/38). pin. Dual-entity dated paint plan recovered. Not n=1540, not a Mem0 same-pin. |
| P28 instrument leftover covering + enumerate items S0 this-VM (2026-08-25) | Same store, hybrid on. Instrument-purpose leftover covering prefers purpose lines over ownership/help floods; enumerate items copy that covering sentence. 113/180 (MH 18/33, OD 4/11, SH 61/98, temporal 30/38). pin. Smartwatch reminder recovered. P27 112/180 hold is not a pin. Not n=1540, not a Mem0 same-pin. |
| P30 what-made leftover covering + lexical admit S0 this-VM (2026-08-25) | Same store, hybrid on. What-made leftover covering prefers off-query evidence over enjoy/participate restatement; lexical search drops made/part, the queried person, and short reason verbs so first-person cause lines enter the packet. 114/180 (MH 18/33, OD 4/11, SH 62/98, temporal 30/38). pin. Running-group push recovered. P29 111/180 dip is not a pin. Not n=1540, not a Mem0 same-pin. |
| P31 how-describe lexical admit S0 this-VM (2026-08-25) | Same store, hybrid on. How-does/did/do-X-describe-Y lexical search drops describe*, capitalized person tokens when other object tokens remain, and got after person filtering so first-person leftover lines enter the packet. 115/180 (MH 18/33, OD 4/11, SH 63/98, temporal 30/38). pin. Stuffed-animal good vibes recovered. First P31 180 114/180 flake is not a pin. Not n=1540, not a Mem0 same-pin. |
| P32 host leftover covering + session admit S0 this-VM (2026-08-25) | Same store, hybrid on. What-did/does/has-X-host leftover covering prefers hosted-event lines; bounded session-id fetch for leftover-covering candidate sessions; hosted-event rank boost; host queries keep leftover hosted-event episodes and join two covering lines. 116/180 (MH 18/33, OD 4/11, SH 64/98, temporal 30/38). pin. Veterans party + share-stories recovered. First P32 180 115/180 hold is not a pin. Not n=1540, not a Mem0 same-pin. |
| P33 advice leftover covering + session admit S0 this-VM (2026-08-25) | Same store, hybrid on. What-advice leftover covering prefers hortative / first-person-gerund directive lines; bounded session-id fetch for advice-echo sessions; search floor for leftover with no query tokens; advice queries keep leftover hortative episodes and join up to three directive lines. 117/180 (MH 18/33, OD 4/11, SH 65/98, temporal 30/38). pin. Gina advice recovered. Not n=1540, not a Mem0 same-pin. |
| P34 what-kind like-list leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-kind leftover covering prefers like-A,-B,-and-C leftover; bounded session-id fetch ignoring spread/kind restatement tokens; search floor for leftover with no query tokens; what-kind queries keep leftover like-list episodes; crowded hop-dump skip excepts kind-list leftover only on what-kind queries. 118/180 (MH 18/33, OD 4/11, SH 66/98, temporal 30/38). pin. Dinner-spread like-list recovered. First P34 180 116/180 dip is not a pin. Not n=1540, not a Mem0 same-pin. |
| P35 how-describe-process prefix hortative leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-describe-process leftover covering prefers prefix hortative leftover (just keep) over companion slogans; sentence-initial just is hortative; process covering empty unless hortative and not process-restatement; hortative leftover sessions seed neighbors first. 119/180 (MH 18/33, OD 4/11, SH 67/98, temporal 30/38). pin. Turtle-care recovered. Does not drop turtles/care. Does not steal Calvin electronic. Not n=1540, not a Mem0 same-pin. |
| P36 what-motivates first-person object-cause leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-motivates leftover covering prefers first-person object-cause leftover over turtle/have-faith/occupation companions; lexical search drops motivate*/keep/even only on what motivates queries; skip hops and hybrid when the search packet already has a cause leftover. 120/180 (MH 18/33, OD 4/11, SH 68/98, temporal 30/38). pin. Joanna writing recovered. Does not drop motivate globally. Does not treat stay-motivated as what-motivates. Not n=1540, not a Mem0 same-pin. |
| P37 what-say-about they-evaluative leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-say-about leftover covering prefers short they-evaluative leftover over dance-photo flood; lexical search drops say*/about only on what does/did/do … say about queries; leftover-covering session fetch is 200 rows; enumerate items copy that covering sentence. 121/180 (MH 18/33, OD 4/11, SH 69/98, temporal 30/38). pin. Dancers graceful recovered. First P37 180 120/180 hold is not a pin. Does not drop say globally. Does not add a dance dictionary. Not n=1540, not a Mem0 same-pin. |
| P38 what-say-about first-person got leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-say-about leftover covering admits first-person possessive/abundance leftover (it's got + quantifier) when another packet line covers the object tokens; they-evaluative leftovers still win; lexical search drops question-frame participles and treats all-caps tokens as non-people. 122/180 (MH 18/33, OD 4/11, SH 70/98, temporal 30/38). pin. NYC It's got so much to check out recovered. Does not treat it's gotta as it's got. Does not steal Tim injury doctor-said. Does not add an NYC dictionary. Not n=1540, not a Mem0 same-pin. |
| P39 dated reported-speech leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-say-about leftover covering admits dated reported-speech leftover (the <role> said + copula) when the query names a calendar date; they-evaluative leftovers still win; undated queries still prefer first-person got leftover. 123/180 (MH 18/33, OD 4/11, SH 71/98, temporal 30/38). pin. Tim injury The doctor said it's not too serious recovered. Does not add a doctor/injury dictionary. Does not reuse they-copula. Does not steal NYC It's got. Not n=1540, not a Mem0 same-pin. |
| P40 how-react leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-react leftover covering admits object-linked they-were observation leftover on how do/does/did … react/respond to; covering requires the observation line itself to name a query object; session expand seeds from FTS hits; list evidence-set cap re-inserts those observations. 124/180 (MH 18/33, OD 4/11, SH 72/98, temporal 30/38). pin. Dogs/snow they were confused recovered. Does not drop react globally. Does not add a dog/snow dictionary. Not n=1540, not a Mem0 same-pin. |
| P41 what-did-purpose leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-did-purpose leftover covering admits adjacent purpose-pair action leftover on what did/does/has … do … to …; covering requires first-person or named query actor; lexical search drops month/year/comparatives only on that query shape. 125/180 (MH 18/33, OD 4/11, SH 73/98, temporal 30/38). pin. Dog-owners group recovered. Does not add a dog/group dictionary. Does not drop November globally. Not n=1540, not a Mem0 same-pin. |
| P43 how-long-been leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-long-been leftover covering admits continuing-years leftover on how long + been; covering prefers duration is N years / N years already / for N years over copula status; lexical search drops long only on that query shape; session rank prefers duration leftover sessions over recency chatter. 127/180 (MH 18/33, OD 4/11, SH 75/98, temporal 30/38). pin. Married 5-year duration recovered. Does not add a marriage dictionary. Does not drop long globally. Not n=1540, not a Mem0 same-pin. |
| P44 how-often leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-often leftover covering admits cadence leftover on how often + does/did/do; covering requires ≥2 remaining object tokens so road-trip months lose to meetup once a week; lexical search drops often and trailing for adjuncts only on that query shape; session rank prefers cadence leftover sessions over recency park-playdates. 128/180 (MH 18/33, OD 4/11, SH 76/98, temporal 30/38). pin. Meetup once-a-week recovered. Does not add a how-often/playdate dictionary. Does not drop often globally. Not n=1540, not a Mem0 same-pin. |
| P45 currently-working leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-project leftover covering admits currently-working leftover on what + project + working/work; covering prefers currently working / working on a new over childhood desire or creating-own restatement; lexical search drops trailing in … course adjuncts and project only on that query shape; session rank prefers currently-working leftover sessions over recency comic-sketch chatter. 129/180 (MH 18/33, OD 4/11, SH 77/98, temporal 30/38). pin. Football-simulator currently-working leftover recovered. Does not add a football/comic-sketch dictionary. Does not drop game globally. Not n=1540, not a Mem0 same-pin. |
| P46 become-interested leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. Dated what-new-hobby leftover covering admits first-person become-interested leftover on what + hobby + interested/become plus a calendar date; covering prefers become interested over a foreign-person new hobby restatement; lexical search drops hobby only on that query shape; session rank prefers become-interested leftover sessions over recency metal-detecting chatter. 130/180 (MH 18/33, OD 4/11, SH 78/98, temporal 30/38). pin. James extreme-sports leftover recovered. Does not add an extreme-sports dictionary. Does not steal John's metal detecting. Not n=1540, not a Mem0 same-pin. |
| P47 how-plan-dream leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-plan-dream leftover covering admits first-person gathering/watching/guide leftover on how does/did/do/will + plan/pursue + dream + learn; covering prefers gathering information / watching videos / beginners' guide over a foreign-person learning their stories restatement; lexical search drops plan/pursue/dream/learning only on that query shape; session rank prefers prep-plan leftover sessions over recency exploring chatter. 131/180 (MH 18/33, OD 4/11, SH 79/98, temporal 30/38). pin. Jolene surf-plan leftover recovered. Does not add a surf dictionary. Does not steal Deborah exploring. Not n=1540, not a Mem0 same-pin. |
| P48 focusing-besides leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. Focusing-besides leftover covering admits first-person focusing-on leftover that covers the besides-object and joins a possessive conjunct on what has/have/is/was + focus/focusing + besides/except/aside; covering prefers focusing on + and my / and our over occupation leftover that also says focusing on; lexical search drops besides/except/aside/lately only on that query shape; session rank prefers possessive-join leftover sessions over recency occupation chatter. 132/180 (MH 18/33, OD 4/11, SH 80/98, temporal 30/38). pin. Jolene relationship leftover recovered. Does not add a relationship dictionary. Does not steal engineering leftover. Not n=1540, not a Mem0 same-pin. |
| P49 titled-show leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-new-series leftover covering admits quoted titled-show leftover on what + new + series/show; covering prefers a quoted watch/show/titled leftover over generic excited about this new journey leftover and quoted novels; lexical search drops new/fantasy/tv/series/excited/about only on that query shape; subject corpus lists 400 rows so titled leftover at recency 314 can seed session expand. 133/180 (MH 18/33, OD 4/11, SH 81/98, temporal 30/38). pin. Wheel of Time leftover recovered. Does not add a Wheel of Time / fantasy dictionary. Does not steal Name of the Wind or Game of Thrones. Not n=1540, not a Mem0 same-pin. |
| P50 locative-purpose leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. Recently-at leftover covering admits first-person locative leftover that names an extra for my/for our purpose object on what did + do + recently/lately + locative at/in; covering prefers that leftover over thin dated locative compiler facts; lexical search drops recently/lately only on that query shape; fact-primary recall keeps locative-purpose episode leftover. 134/180 (MH 18/33, OD 4/11, SH 82/98, temporal 30/38). pin. Japanese-house album leftover recovered. Does not add an album dictionary. Does not steal Dave congratulations leftover. Not n=1540, not a Mem0 same-pin. |
| P51 experiencing-feeling leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-feel leftover covering admits first-person experiencing leftover that names a new level of feeling on how + about + feel/felt; covering prefers that leftover over process restatements of practicing mindfulness and thin experiencing compiler facts; lexical search drops feel/about/progress only on that query shape and keeps mindfulness/gratitude; fact-primary recall keeps experiencing-feeling episode leftover. 135/180 (MH 18/33, OD 4/11, SH 83/98, temporal 30/38). pin. Mindfulness-joy leftover recovered. Does not add a joy/happiness/mindfulness dictionary. Does not steal Deborah mix-of-happiness leftover. Not n=1540, not a Mem0 same-pin. |
| P52 coordinated-use leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-do leftover covering admits first-person-plural leftover that names hard work and determination on what do/does/did + two named hop entities + use; covering prefers hard work + determination + we/us/our over dedication restatements, Calvin-only compiler facts, and car-restoration determination leftover; lexical search drops use/reach/goal only on that query shape and keeps person tokens; fact-primary recall keeps work-determination episode leftover. 136/180 (MH 18/33, OD 4/11, SH 84/98, temporal 30/38). pin. Hard-work-and-determination leftover recovered. Does not add a hard-work / goals dictionary. Does not steal dedication leftover as Calvin-only. Not n=1540, not a Mem0 same-pin. |
| P53 self-directed realize leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. What-did leftover covering admits first-person leftover that names a self-directed realize on what did/does/has + realize + after; covering prefers realize + self-/myself + actor over others-directed support leftover, thin believes-self-care compiler facts, and charity-race attendance leftover; lexical search drops realize/after only on that query shape and keeps event tokens; fact-primary recall keeps self-directed realize episode leftover. 137/180 (MH 18/33, OD 4/11, SH 85/98, temporal 30/38). pin. Charity-race self-care leftover recovered. Does not add a charity-race or self-care dictionary. Does not invent Sunday. Does not drop realize globally. Not n=1540, not a Mem0 same-pin. |
| Honesty stop (2026-08-25) | Isolated leftover covering on this skip-ingest 180 is saturating. No more covering from this 180's leftover ledger. Do not treat 137/180 as 90%, n=1540, or a Mem0 same-pin. Industry on this tenant is still 62/180. Full product /recall is still 11.4%. audit. |
| P54 entity-scoped how-many S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: hop the counted predicate; entity-scope how-many; collapse bare class labels; keep quantified/modified class phrases; specific heads intersect the typed set. 140/180 (MH 20/33, OD 4/11, SH 85/98, temporal 31/38). pin. Children 7→3, ankle 38→2, Sep pets 13→1. Unique losses none. Dec pets still 4 vs gold 3 (Scout stored as Andrew). Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P58 historical typed-set lists S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: scan past current-state for item/skill/place-set/activities+done questions; singular location stays a point fact; leftover how/why does not widen hop scan. 141/180 (MH 21/33, OD 4/11, SH 85/98, temporal 31/38). pin. John outdoor hiking+mountaineering. Unique losses none vs P54. Collars/tricks still miss. Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P59 dest-class lists S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: dest-being skills without requiring trick; transfer-cue item objects (buy/for); possessed-class dest names (dog named, not generic named). 143/180 (MH 23/33, OD 4/11, SH 85/98, temporal 31/38). pin. Audrey collars/tags/toys/beds; James swim/frisbee/skateboard. Unique losses none vs P58. p59/p59b 142/180 are not pins. Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P61 beneficiary org sets S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: who/which beneficiary questions recover raise/for-cue affiliation objects (shelter, homeless, hospital) instead of leftover tournament slogans. 144/180 (MH 24/33, OD 4/11, SH 85/98, temporal 31/38). pin. Unique losses none vs P59. p60/p60b 143/180 food-set 180s are not pins. Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P62 community participation sets S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: in-what-ways community questions recover joined/organizing/host/article-participating-in objects (activist group, art show, mentorship, pride event) instead of leftover courage slogans; attended compiler junk is ranked after. 145/180 (MH 25/33, OD 4/11, SH 85/98, temporal 31/38). pin. Unique losses none vs P61. Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P63 polar has-tried from love S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: has-tried polar hops preference; love/discovered-love/tried of the named person's claim activity proves Yes and locks hybrid; plan/learn/guide and joined hop-Value dumps do not. 146/180 (MH 26/33, OD 4/11, SH 85/98, temporal 31/38). pin. Unique losses none vs P62. Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P64 journey-change historical facets S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: what/which + change + journey recover lists superseded current-state facets (singleton relationship_status hid earlier journey slots) and replaces identity dumps when two+ slots land. 147/180 (MH 27/33, OD 4/11, SH 85/98, temporal 31/38). pin. Unique losses none vs P63. Unpinned who-told 180 was 146/180 unique 0/0 (do not invent Sam). Not leftover covering. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P65 caption observed_at when-event dating S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: when-event dates score each hop memory from its own content (sibling health slots do not share tokens); focus search lets untyped caption ObservedAt compete past recency-400; hybrid abstain does not wipe a typed date. 148/180 (MH 28/33, OD 4/11, SH 85/98, temporal 31/38). pin. Unique gain conv-43-q48 16 November 2023. Unique losses none vs P64. Not leftover covering. Not an ankle dictionary. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P66 entity-scoped language ranking S0 this-VM (2026-08-26) | Same store, hybrid on. S2 product: which/what language+learn/study ranks matrix “is learning X” over purpose adjuncts and other-person slots; one-token objects must appear capitalized next to a learn/study verb. 149/180 (MH 28/33, OD 4/11, SH 86/98, temporal 31/38). pin. Unique gain conv-43-q163 German. Unique losses none vs P65. date_focus leftover-covering skip 180 was 143/180 unique 1/6 (not a pin). Not leftover covering. Not a language dictionary. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P81 titled-work + how-feel-when covering compact S0 this-VM (2026-08-28) | Same store, hybrid on. Class-scoped favorite titled-work ranking plus leftover covering constraints (kinship/find-object/how-feel-when, relative weekend-before, locative caption compact, blessing leftover compacted to a feeling restatement). 150/180 (MH 28/33, OD 4/11, SH 87/98, temporal 31/38). pin. Unique gain vs P66: conv-48-q172 Monster Hunter: World. Unique leftover losses none. P70–P80 180s were not pins (unique leftover losses). Covering hit-rank stays reverted. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P82 alphanumeric short-token covering + start-year compiler S0 this-VM (2026-08-28) | Same store, hybrid on. Distinctive/covering tokens keep alphanumeric shorts (2d) not calendar days (24); when-event covering skips light activityGerundStop gerunds; compiler emits started … in YEAR / since MONTH YEAR (no extraction version bump; 135 not re-extracted). 151/180 (MH 28/33, OD 4/11, SH 87/98, temporal 32/38). pin. Unique gain vs P81: conv-47-q52 June 2022. Unique leftover losses none. P82a 150/180 unique 1/1 (Jasper) is not this pin. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P84 year-start hedged duration-as-of S0 this-VM (2026-08-28) | Same store, hybrid on. Which-year+start covering skips dated snapshots and rewrites hedged for about N years as of {date} (ISO dates parse) to start year; compiler binds doing-them/it to a named activity (no extraction version bump; 135 not re-extracted). 152/180 (MH 28/33, OD 4/11, SH 87/98, temporal 33/38). pin. Unique gain vs P82: conv-48-q50 2020. Unique leftover losses none. P83 151/180 unique 0/0 (currently-playing) is not a pin. Not n=1540, not a Mem0 same-pin. Not 90% (162/180 on this sample; n=1540 for public LoCoMo). |
| P42 how-did-start leftover covering S0 this-VM (2026-08-25) | Same store, hybrid on. How-did-start leftover covering admits duration-matched inception leftover on how did … start … years ago; covering prefers a multi-stem inception pair over a walking-only duration fact or a gym transformation/journey restatement; lexical search drops start/journey wrappers only on that query shape. 126/180 (MH 18/33, OD 4/11, SH 74/98, temporal 30/38). pin. Diet+walking pair leftover recovered. Does not add a diet/walking or gym dictionary. Not n=1540, not a Mem0 same-pin. |
| BEAM 100K this cycle | beam-100k-fresh-20260815.md |
| BEAM 100K historical | beam-100k-c0-async.md |
Historical staging smoke (2026-07, do not mix with R4h): locomo-smoke.md.
Research notes: docs/research.