Anders Swanson distinguishes durable memoryAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts. from a context windowA context window is the maximum amount of tokenized information an AI model can consider during one processing session., temporary files and ad hoc retrieval. Anders Swanson presents memory as a lifecycle covering formation, recall, evolution, evaluation and governance, with different roles for episodicEpisodic memory in AI stores records of specific past interactions or events so a system can retrieve experience associated with a particular situation., semantic, procedural and working information.
Anders Swanson describes combining text, vectors, graphs, timestamps and audit records, then retrieving selectivelyRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task. within a user's authorised scope. Redacting sensitive information before enrichment and enforcing tenancy at both read and write boundaries are part of the design, not optional checks added after memory is stored.
Anders Swanson recommends compact context cards carrying provenanceData provenance records where information came from, how it changed, and which people, systems, or processes handled it. and feedback rather than indiscriminately loading everything. Examples of stale preferences, poisoned information, excessive recall and silent drift show why memory needs explicit tests and revision policies; storage alone does not make an agent trustworthy.
Watch on YouTube
