What happens when AI characters can remember?

Because then everyone knows everything the instant it happens, and the story dies. People — and agents, users, and accounts — only know what actually reached them. So the hard part is not storing memories; it is deciding who is allowed to have which one, and being able to show later how it got there.

New to Memseek? You write one small configuration file describing what your application should remember. After that your application only appends what happened; Memseek does the deriving, keeps every conclusion linked to the evidence underneath it, and hands your agent a bounded briefing instead of a pile of text. Everything below is that file and what it produces — how it works.

This is a useful engineering approximation, not a claim that human memory can be reduced to three numbers. In Memseek, the weights are visible configuration, so a developer can inspect and change what the software favors.

Because each saved statement names the earlier memory it came from, the system can later trace a belief back through the conversations that carried it. That is what “cited memory” and “provenance” mean on this page: the software keeps the receipts.

# 1 · Klaus relays a memory. His statement points to his own source.
await memseek.records.ingest(
    collection="main", type="chat", entity="agent:klaus",
    text="Klaus told Sam that Isabella is planning a party at Hobbs Cafe.",
    content={"heard_from": "agent:isabella"},
    derived_from=[klaus_source_record_id],   # ← the memory Klaus recalled
)

# 2 · about to act: recall across observations, plans, reflections
recall = await memseek.search(
    query="what should I do this afternoon?",
    entity="agent:klaus", mode="hybrid", k=12,
    include=["text", "scores", "occurred_at"],
)
for hit in recall["hits"]:
    hit["scores"]["importance"]   # the paper's term, on the record

# 3 · assemble the prompt: summary + schedule + relevant memory
prompt = await memseek.render_artifact(
    "daily_agent_prompt", entity="agent:klaus",
)
prompt["rendered"]                        # hand this to your model
prompt["manifest"]["input_record_ids"]   # exactly what it knew

# 4 · your simulator's own call — memseek is not in this loop
line = await your_model.say(prompt["rendered"], in_character="klaus")

The runnable example is examples/generative_agents_toy.py plus the shipped Memseek catalog. Dialogue and reflections are generated by a real model at run time, so the wording on this page is explanatory rather than a captured transcript. The architecture is adapted from Park et al., Generative Agents: Interactive Simulacra of Human Behavior, UIST '23.