Akhrots/LAM8B

LAM — Large Akhrots Model (8B)

LAM is a long-term memory model by Tronocity Labs. It does two jobs:

• Store — decides what's worth remembering from a conversation and how it should enter memory (insert / update / merge / skip).
• Recall — answers a question from the user's stored memories, combining and time-ordering them when needed — and says "I don't have that in your memory." when the answer genuinely isn't there, instead of making something up.

Grounded refusal instead of hallucination is the point of a dedicated memory model.

Benchmarks

Scored on the model's answering ability given retrieved memories (token-F1):

Benchmark	LAM
LOCOMO (394)	0.66
LongMemEval-reserved (150, held out)	0.48
LOCOMO abstention F1	0.85

LAM roughly doubles LOCOMO answer quality over its un-tuned starting point and sharply reduces over-refusal.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("atharvalepse0129/LAM8B")
model = AutoModelForCausalLM.from_pretrained(
    "atharvalepse0129/LAM8B", trust_remote_code=True, device_map="auto")

trust_remote_code=True is required — LAM registers its own model class.

Intended use

LAM sits behind a retrieval pipeline: your system retrieves candidate memories and LAM reads them to answer or abstain. It is not a general chat model.

Limitations

• Scores measure answering given retrieved memories, not the retrieval step — end-to-end quality also depends on your retriever.
• Multi-hop (chained) reasoning is the weakest ability.
• This release was trained at 1024-token context.

License

Apache-2.0. LAM is built by fine-tuning open, Apache-2.0-licensed model weights; the derivative is released under the same license.