Agent Memory

Overview

The page begins with flat retrieval-augmented generation (RAG) as a conceptual comparison: a query yields similar snippets, but similarity alone does not make a record current, trustworthy, permitted, or traceable. The working chapters then follow the prototype's captured data rather than recreating a policy or ranker in TypeScript.

Memory model

Memory OS v0 treats memory as evidence under governance rather than a single vector index. A correction supersedes an earlier preference by closing its validity window; the original remains auditable without answering later queries. A session-scoped instruction can remain local to its session, and sensitivity or trust can prevent a relevant-looking record from entering context.

Trace

The trace contrasts a poisoned web observation with instruction-shaped text from a verified private policy actor. The former is neutralized and quarantined as evidence, while the latter is accepted with a recorded policy_authorized decision. A later retrieval example reaches a stale-document episode through a shared entity and names why that traversal entered the cited context packet.

Scope

The demo is not a production agent or a free-form chat interface. Memory OS v0 uses deterministic lexical candidate finding and a one-hop temporal entity index. It does not yet use embeddings, a vector database, background consolidation, or a learned policy.

See also

• arXiv Semantic Graph
• David's Internet

References

1. ^ Agent_Memory/experiments/memory_os_v0, deterministic prototype and trace exporter.