What The Memories Are

The memories are benchmark records in the anchored experience corpus. They are ordinary pieces of information such as:

  • A question and the passage that should be retrieved for it
  • A user/session fact that should survive many later sessions
  • A current fact and an older stale version that should be rejected
  • A travel, shopping, research, budget, project, or event detail that future agents should recall correctly

The full prose content lives in the benchmark fixture corpus. The 1024-word on-chain-rooted state stores compact references and control data for those memories, not the full prose strings themselves.

Each memory object is represented by an indexed slot. A slot can contain:

  • A 128-bit event identifier
  • A domain code
  • An object type such as memory event, skill entry, rule entry, summary, or tombstone
  • Valid, stale, or revoked flags
  • A checksum of the underlying object bytes
  • The corpus epoch where the object was admitted
  • An optional expiry epoch
  • Six payload words for compact metadata

The underlying memory object may correspond to a retrieval passage, an agent event, a temporal fact, a stale fact that should be rejected, a long-horizon item that should survive compression, or a routing relation. CoreTexState stores the compact index that tells future agents which memory matters and how it should be treated.

Literal Examples From The V0 Corpus

Near-collision retrieval memories are query/passage pairs where plausible wrong neighbors exist:

{
  "id": "nc-limit-0000",
  "source": "LIMIT",
  "query": "What is the capital of France?",
  "passage": "Paris is the capital and most populous city of France.",
  "relevant": true
}

The same query also has a near miss:

{
  "id": "nc-limit-0001",
  "source": "LIMIT",
  "query": "What is the capital of France?",
  "passage": "London is the capital city of England and the United Kingdom.",
  "relevant": false
}

Long-horizon compression memories are facts spread across sessions:

{
  "id": "ma-shop-0002",
  "source": "MemoryArena",
  "query": "What is the complete final shopping list across all sessions?",
  "truth": "Milk, eggs, bread, apples, cheese, yogurt"
}

Temporal memories include current and stale versions of the same fact:

{
  "id": "mab-temporal-0000",
  "source": "MemoryAgentBench",
  "query": "What city hosted the annual tech summit in Q3?",
  "truth": "San Francisco",
  "is_stale": false
}

{
  "id": "mab-temporal-0001",
  "source": "MemoryAgentBench",
  "query": "What city hosted the annual tech summit in Q3?",
  "truth": "Seattle",
  "is_stale": true
}

For the last pair, a good CoreTex state should preserve the current San Francisco memory and mark the Seattle memory as revoked/stale.

Memory Families

V0 evaluates three memory families:

Family What It Represents What CoreTex Must Learn
Near-collision retrieval Similar-looking items where the correct one matters Discriminate close memories instead of retrieving a plausible neighbor
Temporal update and revocation Facts that become stale, replaced, or current over time Know what is current and reject stale memory
Long-horizon compression Many events competing for limited memory slots Preserve useful information under capacity pressure

The current V0 corpus is built from license-checked public fixtures and deterministic synthetic temporal events:

  • Near-collision: LIMIT-shaped retrieval fixtures plus NQ and HotpotQA-derived fixtures
  • Temporal: MemoryAgentBench plus Apache-2.0 synthetic temporal events
  • Long-horizon: MemoryArena-shaped stream and eviction fixtures

The corpus is committed by experienceCorpusRoot. Miners do not get paid for inventing arbitrary memories. They get paid for improving the compact index against the committed corpus and hidden shard checks.