Skip to content

agent-memory

A lightweight, vendor-neutral AI memory + cognitive-loop system. No-code, markdown-only. One shared memory/ layer that any AI agent — Claude, Gemini, Cursor, Kiro, GitHub Copilot, Antigravity — can read, write, and evolve.

Get started Read the whitepaper View the deck

Why it exists

AI agents are moving from single-turn prompting to persistent, memory-driven runtimes. Two problems dominate production use:

  • Memory drift & vendor silos

    Context is re-explained every session, decisions are silently forgotten or contradicted, and each AI tool hoards its own steering files. A team using more than one vendor has no shared, durable project memory.

  • The intent ↔ delivery gap

    Agents are creative, but keeping what is built faithful to what was intended is largely unmanaged. "Confidently wrong" facts and reversed decisions accumulate.

agent-memory addresses both with a no-code, markdown-only system layered over a single, git-committed memory/ directory. The files are the product; the agent is the runtime.

Two layers

  • Backward — Evolving Memory

    An event-sourced ledger (immutable session logs) projected into a live continuity.md, with deterministic decay, supersession, invariant re-verification, contradiction checks, and provenance.

    "Where are we, and why?"

    Evolving Memory

  • Forward — the VBDI Loop

    A lightweight cognitive loop — Current State → Vision → Blueprint → Design → Implementation → Feedback — with an enforceable intent trace and a human gate at every altitude change.

    "Where are we going, and is delivery faithful?"

    The VBDI Loop

What makes it different

  • Vendor-neutral. One shared memory; thin per-vendor pointers route every agent to a single hub. No lock-in.
  • Deterministic — no floating-point. Every decision reduces to counting or comparing integers, so results are reproducible across agents and runs.
  • Mechanize the arithmetic, not the judgment. The deterministic parts of a memory ritual (re-tiering, the archive move, adapter sync, integrity checks) are runnable helpers; every act of judgment stays with the agent and the human.
  • No manual user step. Triggers, init, and adapter sync are agent-activated, with CI as the zero-config backstop.
  • Additive, non-destructive upgrades. Versioned; an older repo upgrades in place via an idempotent ladder. Nothing is ever deleted — faded facts archive to a greppable index.

Get going

  • Getting Started

    Install, enable a repo, and run your first session.

    Getting Started

  • Concepts

    How evolving memory, the VBDI loop, skills, and decay work.

    Concepts

  • Guides

    Enable, upgrade, review, resolve conflicts, author skills.

    Guides

  • Reference

    Built-in skills, decay parameters, vendor support, protocol files.

    Reference

Status

agent-memory is independent research published under Apache-2.0. It is not an official Accenture product. The current release is documented against v4.26.1.