Memory architecture for AI agents
An agent's memory that lives entirely in text files.
No database. No vectors. No proprietary model. Vibrating Memory is an architecture of memory cards — facts, index, portrait, instant perceptions — designed to run on any popular LLM capable of following rules written in natural language.
Three deliberate constraints
What makes the system reproducible by anyone
01 — storage
No database engine
Markdown for cards, JSONL for logs, raw logs for technical monitoring. Nothing to install, nothing to maintain, nothing to run.
02 — model
Popular LLMs
No exotic capability required. If a model can read rules and follow them consistently, it can carry the system.
03 — logic
Natural language, no code
The cards are text an LLM understands directly. Code stays at the edges — reading a file, handling an API call — never inside the memory logic itself.
A documented origin
Traced priority, not just claimed
-
March 2026
Early experimentation. Prefix-based context caching on live conversational video avatars (LiveKit, Groq) — the base intuition: separate what almost never changes from what changes every turn.
-
August 26, 2026
Formalization of the unified concept. The architecture is named Vibrating Memory. It groups memory by function, not by content type: Knowledge (persona, chats, synthesis, project memory, portrait, people, tools, prohibitions, deliverables, documentation), Cognition (rules, index, skills, network, hot memory, curiosity, tasks, inbox), Sensation (sensors, self-awareness, sensitivity, temporal), Reflection (patterns, private garden, self-understanding), and Modulation (constraints, mood, originality) — what each piece of memory actually does for the agent, not what kind of fact it holds.
-
Ongoing
Public documentation. Progressive publication of the Vibrating Memory series — architecture, factual cards, sensory memory.
Reference document · not a blog post
Vibrating Memory — A Declaration
The complete, dated statement of the whole system in one document — every card, briefly but concretely. Read this first.
Journal
The series, as it's published
-
August 26, 2026 · Part 1 · Published
Vibrating Memory, Part 1 — Giving a Claude Agent Real Memory, Without a Database
The foundation of the system: A/B/I/S/C/E, the Claude API cache mechanism, and the three constraints that make the architecture reproducible by anyone.
-
Coming soon · Part 2 · In progress
Factual cards — portrait, self-awareness, other agents, tasks and skills
Beyond the cache: how the agent builds a lasting understanding of who it serves and how it works.
-
Coming soon · Part 3 · In progress
Sensory and temporal memory
The system's most experimental piece — instant perceptions rather than accumulated facts.