Sample brain
8 memories across a week. In a real workspace, each one was a Slack message, PR, Linear issue, or meeting transcript that the classifier kept (and 5–10× that got dropped as noise).
- decisionslack · #eng-platform
Adopt Postgres + pgvector over Pinecone for vector search
Eng team decided to migrate the embedding pipeline off Pinecone and onto Postgres + pgvector. Reasoning: we already run Postgres for primary data, Pinecone costs 4x more at our scale, and pgvector recall is within 2% of Pinecone on our gold set. We accept ~30ms higher p99 latency in exchange. Revisit if recall drops below 80% or if we exceed 5M vectors.
#architecture#vector-search#postgres - observationlinear · GROWTH-142
Demo conversion drops 18% when we show pricing on the landing
A/B test results (n=2,341 over 14 days): landing page with pricing visible converted at 4.1%; without pricing at 5.0%. Statistically significant (p=0.012). Hypothesis: B2B buyers want to qualify themselves *before* seeing a number. Decision pending — see linked Q3 planning doc.
#growth#pricing#ab-test - learninggithub · frozo-ai/edgegate#482
Always log 5xx with caller IP — debugging silent failures otherwise impossible
Spent 6 hours chasing a 'webhook delivery randomly fails' bug. Root cause: a downstream proxy was rate-limiting one specific IP range. Our 5xx logs only had the request ID, not the originating IP. Lesson: every 5xx path should record `request.ip`, `request.user_agent`, `request.country` at minimum. Filed PR #482 to add this to the standard error middleware.
#debugging#observability#5xx - todoslack · #eng-platform
Ship PII redaction layer before first design partner
Before onboarding the Mumbai design partner, every body string sent to the classifier must pass through a regex sweep for emails/phones/SSN/PAN/Aadhaar/credit-card patterns. Owner: Ashish. Definition of done: tests covering all 8 patterns + classifier prompts changed in all 4 pollable connectors. Target: 2026-05-20.
#security#dpdp#p0 - questionnotion · Q3 planning / GTM
How many design partners need on-prem vs cloud?
Of the 12 enterprises in active outbound, 4 have asked about self-hosting. Two of those are banks (regulatory). One is a defense contractor. One is just paranoid. Need to decide: ship a self-host installer in v1 (slows everything down) vs cloud-only with BYOK + Bedrock (faster ship, narrower TAM). Not resolved.
#gtm#enterprise#self-host - decisionmeeting · Founders sync 2026-05-13
Apply to YC S26 with vault-mem as the lead bet
Frozo and kincare go to maintenance. Vault-mem is the YC S26 application. Reasons: agent-memory category is heating up (mem0, MemGPT, Letta), we have a working OSS + dogfood story that they don't, and the multi-product story (OSS personal brain + Cloud company brain) is uniquely sellable in YC's deck format. Application due 2026-09-30.
#yc#strategy#company - entitynotion · Vendors / LLM
Anthropic — primary LLM vendor
Anthropic provides our classifier (Haiku) and reasoning (Sonnet) models. Commercial API has a no-training commitment in their TOS. SOC 2 Type II, HIPAA BAA on enterprise plans. India region available via Bedrock (relevant for DPDP residency). Direct relationship is preferred over OpenRouter for privacy-conscious tenants — see BYOK docs.
#vendors#llm#compliance - summarygithub · frozo-ai/frozo-vault-mem (commit 0ed2371)
DPDP / GDPR erasure cascade landed (Phase 5 complete)
Per-subject erasure now cascades through markdown vault → embeddings → FTS index → audit log in a single transaction. 8 migrations applied (`20260519180000` through `20260519180007`). MCP tool `memory_erase_subject` exposed for agents. Admin UI at `/admin/subjects`. Audit log hashes the subject_id so plaintext never persists. Tested with 200 synthetic subjects. Owed to Mumbai design partner before launch.
#dpdp#gdpr#compliance#phase-5