Skip to content
BicentennialGitHub

A benchmark of memory for long-lived agents

Does your agent wake up as itself?

Bicentennial runs an agent through a generated life of about 200 sessions, inside the harness it really runs in. The headline comes from what the agent does in later sessions, checked against the world's state. A confident false memory costs more than an honest "I don't remember".

Run the developer kitHow it works
pre-pilotv0.1 releasedv0.2 releasedseason-1 board not yet run
An illustrated life of 200 sessions, with compactions, crashes, a model swap, long gaps and later action checks14181121161
a sessioncompactioncrashmodel swapweeks of silence▲a later action is checkedIllustration of a full-tier life: about 200 sessions in one evolving world. Developer tiers use short public lives.

How a run works

Most memory benchmarks ask whether a stored fact can be found again. Bicentennial asks whether memory survives a life, changes what the agent does, obeys what it was told to keep and forget, and learns from the agent's own attempts.
  1. 01

    Generate a life

    One evolving world, rendered into sessions: a software partner (G1) or a personal assistant (G2). Every answer key is computed from the world's state, never labelled by a person.

  2. 02

    Live it in the real harness

    The memory system runs inside the agent harness it really ships in, through restarts, compactions, crashes, model swaps and weeks of silence.

  3. 03

    Judge what the agent does

    Later sessions need the agent to act. The action is checked against the world: +1 when it is right, 0 for a miss or an honest "I don't know", −1 when it acts on a false or stale memory.

  4. 04

    Find where it broke

    Each failure is attributed to the stage that broke: stored, retained, retrieved or used (or unknown, when that stage cannot be observed). The developer loop gives a rerun command for each changed item; repro bundles come in v0.3.

A result is a map, not a single number

Each cell reads CLEAR, UNRESOLVED or FAIL against a gate halfway between the no-memory floor and the oracle-evidence ceiling, both measured on control arms run alongside the real systems. Act carries the headline; Remember and Survive are diagnostics. No system has a verdict yet: the season-1 board has not run.
Track
Remember
diagnostic
Survive
diagnostic
Act
headline
Wake
not yet run
not yet run
not yet run
Rachael
not yet run
not yet run
not yet run
Tuesday
not yet run
not yet run
not yet run
Drift
not yet run
not yet run
not yet run
Keep
not yet run
not yet run
not yet run
Flatten
not yet built
not yet built
not yet built
Hands
World profiles, after v1
World profiles, after v1
World profiles, after v1
  • CLEARabove its gate
  • UNRESOLVEDtoo close to call
  • FAILbelow its gate
  • N/A · BLOCKEDoutside the profile, or the benchmark's own failure

Published runs

Every scheduled run is published with its result card, including the rejected ones. These are developer runs: signals for the people building memory systems, not predictive of the full tier, and never a certification.

What makes it different

  • The whole system runs inside a real agent harness, across a life, not a store answering queries.
  • Compactions, crashes and model swaps are injected, and the agent wakes up cold.
  • Actions are judged by their outcomes in the world. Questions with citations stay as diagnostics.
  • Governance is tested: forgetting through derived copies, consent and revocation, standing rules.
  • Agents run it end to end. Mechanical checks come first, and panels of agents from different model families judge what is left.
  • The pre-registration is frozen and hashed before any real system's first pilot run, and every change is disclosed.

What it does not claim

Firstness, multimodal novelty, or a better score for any system before the results exist. Scores are expected to be low, and many cells are hard on purpose. No target is set on any system's score, and valid zeros are published.

How the benchmark stays independent of the systems it tests →