Systems landscape
Shared by Pavel · Updated 2026-11-12
The systems differ in what they preserve, who or what controls writes, how revisions affect prior records, and how retrieval assembles context.
Representing memory
- MemGPT/Letta treats memory as editable by the agent through tools, with tiers inside and outside the context window; the agent can move information between working context and longer-term storage.
- Mem0 stores natural-language facts and a conversation summary. For each new message pair, it extracts candidate facts and compares them with similar stored memories.
- Graphiti retains raw episodes as a ground-truth stream, then derives entities and facts linked back to the episodes. Its temporal knowledge graph can represent when facts are valid.
- TriMem combines raw dialogue segments linked to their sources with extracted atomic facts and synthesized profiles.
- A-MEM represents each memory as a contextualized note with keywords and tags, linking related notes in a Zettelkasten-like network.
- Generative Agents keeps observations in a natural-language memory stream and adds higher-level reflections that cite the supporting records.
- Voyager stores procedural memory as an expanding library of executable skills; ReasoningBank distills reusable strategies from an agent’s successes and failures.
Updating memory
- Mem0 chooses among ADD, UPDATE, DELETE, and NOOP after extracting candidates and retrieving similar memories. UPDATE and DELETE overwrite older memories when new information contradicts them.
- Graphiti compares new facts with existing graph edges and can invalidate an earlier edge while retaining it, rather than simply replacing the historical record.
- A-MEM can update older note descriptions when new, related notes are added.
- Generative Agents periodically form reflections from observations; Letta also documents background “dreaming,” in which subagents consolidate recent conversations, sometimes with review before changes apply.
- These approaches reflect different write policies: store observations first and derive later, selectively extract facts, or let an agent decide through explicit memory operations. Each trades off retained detail against storage, noise, and update complexity.
Retrieving memory
- MemGPT/Letta exposes memory through agent tools and tiers; in a filesystem baseline, the model searches files with operations such as grep, opening and iterating over results.
- Mem0 searches for similar stored memories before deciding how a new fact changes them; its documented workflow retrieves the 10 most similar memories.
- Graphiti uses the graph of entities and facts, with raw episodes available as provenance; its temporal structure supports questions about what was true at a given time.
- HippoRAG combines an LLM-built knowledge graph with Personalized PageRank to retrieve across linked concepts.
- Retrieval can also be push—the system places selected memory into the prompt before the model runs—or pull, where the agent searches when needed. Time-aware query expansion can narrow retrieval to the period named in a question.
- More memory is not automatically better: summaries can lose details, while feeding whole histories raises cost and can impair performance as context grows. The guide describes verbatim records, extracted facts, and linked hybrid designs as distinct ways to manage that tension.