Build archive
Building Log
Notes, benchmarks, and decisions from Paolo's development.
Last updated September 2, 2026
Competitive field
| Approach | Differentiator | Limitation | |
|---|---|---|---|
| Paolo | Cognitive profiling via AI conversations | 10-dimension fingerprint, taste model, org knowledge layer | Requires AI conversations (not passive) |
| Supermemory | Universal memory API, 12-way LLM ensemble | 98.6% on their benchmark, open source | No cognitive profiling, expensive per-query |
| Littlebird | Desktop screenreading ($11M raise) | Sees everything across all apps | Desktop only, privacy concerns, no profiling |
| claude-subconscious | Background CC agent (Letta/MemGPT) | Agentic — reads code, whispers guidance | CC only, coding only |
| Mem0 | Memory API for LLM apps | Simple API, broad integrations | No profiling, no org layer |
Everyone in this space is building memory as storage: save what happened, retrieve it later. We're building memory as understanding: learn how someone thinks, what they prefer, how they make decisions. The moat isn't the data you store, it's the model of the person you build from it.
The Cognitive Fingerprint
Paolo builds a 10-dimension Bayesian profile from natural conversation. No quizzes, no onboarding flows. The profile emerges from how you interact with AI.
Benchmarks
PaoloBench
Cognitive memory accuracy across 7 categories. Tests what matters for a personal memory layer: preferences, decisions, patterns, temporal reasoning, cognitive profile, entity knowledge.
MemoryDelta
Personalization lift in blind A/B test. Memory wins 70% of comparisons. The question: does memory make the response actually better for this specific person?
Snapshot Latency
Full profile + preferences + context. Warm cache: 0.4s. Down from 12s+ before optimization.
Recent Milestones
Paolo CC Plugin
One-click install for Claude Code users. Automatic memory capture, context injection, zero configuration.
Write-Time Expansion
Pre-compute search aliases when storing memories, eliminating the read-path LLM dependency for faster, more resilient retrieval.
Cosmos Institute grantee
Cosmos Institute grantee, Oxford Philosophy x AI track, for Memory Without Capture, a study of how persistent AI memory affects user autonomy. Read the announcement.
Agent Memory Layer
Automated agents store and share org knowledge through Paolo's multi-tenant infrastructure. Agents share context with each other and with humans in the same org.
Claude Code Hooks
Automatic session capture with 0.13s context injection. Tool actions captured in background, session summaries synced on exit. No manual memory calls needed.
PaoloBench + MemoryDelta
Two novel benchmarks measuring cognitive memory quality and whether memory actually makes AI responses better for a specific person.
Performance Overhaul
Snapshot latency from 12s to 2s. Circuit breaker for API resilience. Accuracy canary monitors retrieval quality every 6 hours.
Privacy & Security Audit
63-test E2E suite. Credential scrubbing, raw message stripping, entity sensitivity classification, OAuth hardening.
Organization Layer
Multi-tenant knowledge sharing with role-based access. Built for The Operator, designed for any team.
Cognitive Engine v2
10-dimension Bayesian fingerprint from 2,774+ observations.
Founder Thoughts
Everyone is building AI memory. Almost nobody is measuring whether it works.
This week we built MemoryDelta, a benchmark that asks a question nobody else seems to be asking: does having memory actually make the AI response better for this specific person? Not "can you retrieve fact X" but "did knowing X change the output in a way that mattered?"
The answer: memory wins 70% of blind A/B comparisons, with a +3.65 personalization lift on a 5-point scale. But accuracy only improves by +0.05 without confidence signals. Memory makes AI more personal, not more correct. That's a crucial distinction that shaped how we think about retrieval quality going forward.
Curated memory vs captured memory
Littlebird raised $11M to build AI that watches your screen. Supermemory has 17K GitHub stars auto-capturing everything. Both are betting that more data equals better understanding.
We're betting the opposite. Paolo stores what the AI chose to remember, not everything it observed. The user is the author of their own memory, not the subject of surveillance. There's something meaningful about the difference between "I told my AI about myself" and "my AI watched everything I did." We think the first one builds trust. The second one might be more complete, but completeness isn't the same as understanding.
Why we don't win on MemoryBench (and why that's fine)
Supermemory's MemoryBench tests raw transcript retrieval: "what did the user say in conversation 7?" Paolo scores 20% on that benchmark. By design. We deliberately strip raw messages for privacy and store summaries and observations instead.
So we built our own benchmark, PaoloBench, testing what we think actually matters: preference recall, decision tracking, cross-session patterns, cognitive profiling, entity knowledge. We score 95.7%. But we also published our MemoryBench score, because intellectual honesty about what you're bad at matters more than only showing the numbers you like.