Passed
Can each person receive a different, grounded answer?
Two synthetic users asked the same question from isolated memory spaces and received different evidence-backed answers, with no cross-user hit.
Personalization 2/2 · cross-user leaks 0Evaluation story · 22 August 2026
We did not begin with a benchmark. We began with Granddad.
“Older people are not a support problem. They hold decades of memory the rest of us rarely know how to ask for.”
— Rach Pradhan, creator of familiar17/17retrieval and access checks
8/8grounded agent answers
5/5language journeys
3/3generated voice-note trials
The question behind the data
Can sophisticated Gemini infrastructure feel like an ordinary family conversation?
A useful elderly-first product has to succeed on more than model quality. It must understand speech and photographs, preserve language and tone, find the right evidence, keep people isolated, and leave sharing under human control.
Every result below used fictional people and memories. Live fixtures were deleted by their exact manifests after inspection.
Five questions, not vanity metrics
What we tried to disprove.
Passed
Can an older relative begin by speaking naturally?
English, Chinese, Malay, Marathi, and Hindi journeys each held one-question pacing and language continuity after the repair guard.
Multilingual journeys 5/5Passed
Does real Telegram-style audio survive the whole media path?
Synthetic speech was generated, converted to OGG/Opus, then understood again by Gemini 3.7 Flash—including corrected details and a Malay memory.
Audio trials 3/3Passed
Can family connection stay separate from archive access?
A connected Granddad and Auntie began with zero access. One explicit share reached only Granddad; revocation returned active access to zero.
Granddad 1 · Auntie 0 · after revoke 0Passed
Does the agent harness fail closed?
The live Gemini loop selected one owner-bound search tool, kept six Durable Object messages in order, and rejected a mismatched owner before persistence.
Tool calls 1 · ordered turns 6 · owner rejection 1ALMA-inspired, human-reviewed
The first memory design was not good enough.
ALMA treats memory behavior as a design that can be searched and evaluated. familiar borrowed that inspectable loop: preserve a semantic baseline, test candidates on multilingual exact evidence, and explicitly select one reviewed policy. Production never runs model-generated memory code or rewrites permissions.
Retrieval designRecall@1 / MRR
Semantic baselineANN similarity only
Balanced recallSelected policy
This small exact-evidence fixture shows why we kept semantic similarity primary but added language-neutral lexical evidence, duplicate collapse, light recency, and source diversity.
Created during the hackathon
The safety boundary became its own package.
The first harness did not fit a consent-sensitive family archive, so Rach extracted the reusable part as standardharness. It is model-agnostic, but familiar uses it around Gemini.
- 1
Reject the wrong owner or delivery channel before loading context.
- 2
Load a bounded per-user conversation and label history as untrusted data.
- 3
Persist the accepted user turn before asking Gemini to reason or use a tool.
- 4
Persist only a non-empty reply; model failure never fabricates an assistant turn.
Checked-in snapshot
The complete evidence, including the limits.
- App + package tests
- 70/70 passed
- Production retrieval and ACL
- 17/17 · Recall@8 1.00 · MRR@8 1.00
- Grounded answers
- 8/8 · invalid citations 0 · forbidden leaks 0
- Gemini Embedding 2
- 768 document values · 768 query values
- standardharness boundary tests
- 5/5 passed
- Current limits
- Small synthetic cohorts, local long polling, native-speaker review still needed
The conclusion
The models can be complex. Remembering together should not be.
familiar gives every generation one private place to speak, ask, and choose which stories travel forward.