memonative · paradox shell · v1.0.0
paradox>working...
// devs: try `sudo` to unlock the memonative api surface
See What I'm Building →

Amay Mani Tripathi

I 

build AI memory systems

• AI Infrastructure • Cognitive Architecture • Neuroscience × AI • Memory Systems
• AI Infrastructure • Cognitive Architecture • Neuroscience × AI • Memory Systems
• Vibe-Coded Engineering • Open Source AI • Entrepreneurship
01focus area

Cognitive Memory Systems

I build AI memory that works like human memory — with biological forgetting, reconsolidation, and trust-weighted retrieval. Not flat key-value stores. Living, evolving memory graphs that strengthen with use and gracefully fade when irrelevant.

02focus area

AI-Powered Research & Intelligence

I design systems where AI agents conduct deep, multi-source research autonomously — synthesising web intelligence into structured, actionable knowledge graphs. Making AI do the homework so humans can make the decisions.

03focus area

Production AI Infrastructure

I ship AI systems that run in production — async event-driven pipelines, background workers, database migrations, comprehensive stress tests. Not Jupyter notebooks. Not API wrappers. Real, deployable, scalable cognitive infrastructure.

04focus area

Vibe-Coded Engineering

I leverage AI-assisted development to build technically sophisticated systems at speed. AI accelerates the implementation — genuine domain expertise in neuroscience and systems engineering drives the architecture and design decisions.

How I Think About AI

"I don't build wrappers around APIs and call them AI products. I build cognitive architectures that solve genuinely hard problems memory systems that forget, research engines that think, infrastructure that actually ships. The future of AI isn't just smarter models. It's smarter memory, smarter tooling, and builders who understand both the neuroscience and the engineering trade-offs."

India · Building in Public · Exploring AI Consciousness
03section

Projects

cognitive infra in production. fewer wrappers, more substrate. the page below dives into one build at a time.

01featured build

Memonative

Functional MVP

A neuroscience-inspired memory engine that forgets, strengthens, corrects, and consolidates — just like the human brain.

//problems it solves
problem.01
the flat memory problem — agents that recall everything equally fail to recall what matters
problem.02
the amnesia problem — long-running sessions drop context the user assumed was kept
problem.03
the lost-history problem — facts get overwritten with no trail of what changed or why
//approach

Three memory types modelled on the established neuroscience taxonomy — episodic, semantic, procedural — each with biologically plausible decay rates and reinforcement dynamics. Ebbinghaus decay curves, memory reconsolidation, episodic-to-semantic consolidation, and trust-weighted retrieval, all behind a JSON-only HTTP surface designed for agent tool-use.

//benchmark
longmemeval-s · recall@5● gpt-4o-mini · stratified n=100
multi-session100%long-haystack recall
temporal-reasoning100%time-ordered facts
single-session-user92.9%in-session callback
knowledge-update100%retraction handling
evaluated on longmemeval-s · stratified slice · pinned to gpt-4o-mini + text-embedding-3-small.
//stack
core
Python 3.11+FastAPIasyncpg
storage
PostgrespgvectorAlembic
workers
CeleryRedis
ai
OpenAI embeddings + LLM
//highlights
01Full revision chain tracks WHY facts changed — retraction vs temporal_update, with timestamps.
02Episodic Consensus Engine turns repeated low-confidence episodes into high-confidence semantic facts.
03Hybrid retrieval — vector + Postgres FTS, fused via reciprocal rank — beats pure-vector on long-haystack.
04Per-tenant row-level isolation with SHA-256 hashed bearer tokens at rest.
Source closed for nowpaper + open-source release queued — reach out for a walkthrough
last_updated · 2026-05
Built by Amay Mani Tripathi · 2026