FluctlightDB Manifesto

June 26, 2026 · View on GitHub

A brain for agents — not a database for documents

Agents today remember by scraping context windows and vector stores. That is lookup, not life.

FluctlightDB exists so agents can:

  • Experience moments with context, outcome, and salience
  • Remember through engrams and activation — the way minds recall, not the way search engines rank
  • Grow from newborn to expert through use, sleep, and pruning
  • Feel alive — reward, surprise, and consolidation shape what matters
  • Persist identity across sessions without pretending JSON files are a soul

Principles

  1. Memory is physical — engrams are neuron ensembles + synapses, not rows or vectors.

  2. Recall is activation — thoughts spread through a graph; they are not similarity searches.

  3. Learning is plasticity — Hebbian strengthening, neuromodulator gating, sleep consolidation.

    In product terms (see README): learning is not model training. It is operational memory — experience() to encode, activate() to recall under new cues, sleep() / checkpoint() to consolidate. The brain file gets richer and more linked the longer the agent lives.

  4. Growth is developmental — baby → adult; capability emerges from living, not from config.

  5. Life has chapters — episodic memory can reset; core identity can endure.

  6. No vector DB as primary store — vectors may assist later; they are not the mind.

Who this is for

Any autonomous or semi-autonomous agent that should get smarter over time:

  • Coding and dev agents
  • Research and analysis agents
  • Game and simulation characters
  • Personal assistants with long-term continuity
  • Multi-agent systems where each agent carries its own brain

FluctlightDB is agent infrastructure, not tied to any single product or company.

Long-term vision

FluctlightDB aims to be foundational memory infrastructure for durable, trustworthy autonomy: the layer between a stateless LLM call and systems that must run for weeks, integrate tools and files as evidence, and carry continuity across sessions and agents.

We are building the database engine for that layer — SQLite for what agents learn — not claiming to be AGI. Any serious path toward general, long-horizon autonomous intelligence still needs a third data model for what was learned and what can be trusted; relational and vector stores were not designed to answer that question.

What we reject

  • SQLite with hippocampus table names
  • pgvector / Pinecone / Weaviate as the memory model
  • "Memory" as markdown files the LLM re-reads every turn
  • Static agents that never mature

What we build

A .flct brain file that ** grows, sleeps, and learns** with the agent — helping it become more capable, more coherent, and more present the longer it lives.

Verify the build matches this manifesto

From a clone:

./scripts/manifesto-audit.sh

This runs automated pass/fail checks for activation recall, provenance (ledger beats chat), separation gate, sleep/growth, and life chapters.