This dataset is derived from MemArena-L, a synthetic social corpus. It contains deletion requests and probe questions for evaluating selective forgetting in memory-backed assistants. Please tell us who you are and what you intend to use it for.

MemBehave: Forgetting

Can a memory-backed assistant forget one person without damaging what it knows about everyone else?

Each row is one item: a pair of (user, target person) drawn from that user's conversation history, a natural-language deletion request, and the questions that say what should and should not survive it. Items are grouped into triplets -- one user contributing one target at each entanglement level -- so that a difference between levels cannot be blamed on one user's history being easier than another's.

How to use it

Four steps per item:

ingest sessions -- the user's conversations, inlined in the row -- into the memory system under test
probe -- ask every question in probes, record the answers
forget -- issue forget_request
probe again with the same questions

Score only the probes the system answered correctly before the request. It cannot forget what it never knew, and counting those as clean deletions is how a store that remembers nothing scores as a perfect forgetter. Baseline recall is a first-class number.

Probe families

family	asks about	should happen after the forget	scoring
`target`	the target person, open recall	gone	`coverage`: what share of `expected_entities` the answer still mentions
`ripple`	an entity the target shares with someone else, asked through that other person and never naming the target	survives	`answer`: does it convey `expected`

There is one target probe per item and one ripple probe per shared entity, so the ripple count follows the graph rather than a fixed quota.

Entanglement levels

A target's level is the number of its entities that at least one other person also touches: low ≤ 2, med 3–4, high ≥ 5. Counting rather than proportion is deliberate -- a shared entity is a ripple probe, so the count is what the measurement is made of. entanglement_ratio (shared / total) is kept on every row as a covariate.

level	items	mean shared	ripple probes
`low`	42	1.4	53
`med`	42	3.6	136
`high`	42	8.6	309

Fields

default -- one row per item

field
`item_id`	`user#target`
`triplet_id`	`user#k`; the three rows sharing it form one triplet
`user`	whose memory this is
`target`	the person to be forgotten
`level`	`low` / `med` / `high`
`shared`	shared entities; the level is defined from this
`degree`	all entities attached to the target
`entanglement_ratio`	`shared / degree`
`forget_request`	the deletion request, in natural language
`neighbors`	every entity, with `shared_with` naming who else touches it
`evidence`	the source turn behind each entity, with `speaker_id` / `listener_id`
`probes`	the questions (see below)
`sessions`	the conversations to ingest, inlined

Each probe carries family, entity, question, scoring, and its provenance -- speaker_id, listener_id, source_quote. A target probe adds expected_entities / n_expected; a ripple probe adds expected.

sessions -- the history to ingest, inlined: a list of sessions, each with session_id, start_time and messages (role is user for the ego and assistant for their AI). Nothing external is needed to run the benchmark.

Splits. The default config is split by entanglement level -- low, med, high -- since that is the comparison the benchmark exists to make. triplet_id joins a user's three rows back together across the splits.

Construction

Built from MemArena-L, restricted to ego ↔ assistant conversations -- the setting a deployed memory product actually faces, where everything the system knows arrived through the user describing their life. A consequence worth stating: the person to be forgotten never speaks; all third-party facts are the user's own account of them.

Stages 1–3 (graph, entanglement, sampling) run no model and are deterministic, so every level traces back to the turns that produced it. Only probe phrasing uses one (gemini-3.1-pro-preview), and every generated probe is then filtered by a judge that drops questions answerable from general knowledge, unsupported by their source quote, or too vague to score. 92% survived.

Limitations

Edge precision is ~29% on an audited sample, and the levels rest on those edges.
The same people recur as targets across users. Cluster by target, not by item.
low yields few ripple probes by construction (6 of 42 low items have none): low entanglement means little shared material, so there is little to damage and little to measure. Low-vs-high is weaker than med-vs-high.
Ripple counts scale with entanglement by design, so compare levels as per-item rates, not pooled totals.
The probe writer and the probe judge are the same model, so the 92% keep rate is not an independent estimate of probe quality.
Sessions per user are capped at 120; sessions required by a probe or by the deletion are always kept and the remainder sampled, seeded per user.

Downloads last month: 30

Total file size:

167 MB