7 sections·~4 min read·1 expanded
For each paper below, read for four things specifically: (1) what exact problem are they solving that prior work didn't, (2) what's the core mechanism/algorithm, (3) what benchmark did they use and what were the actual numbers, (4) what's the stated limitation even the authors admit to. This is worth doing on paper — literally write these four things down for each paper before moving to the next.
- Problem: existing memory approaches lack a systematic, benchmarked comparison; also proprietary systems (like a major AI lab's built-in memory feature) underperform specifically on temporal reasoning due to poor timestamp extraction.
- Mechanism: the ADD/UPDATE/DELETE/NOOP routing pattern you built in Module 5, plus a graph-extended variant (Mem0^g) using Neo4j for relationship storage.
- Benchmark: LoCoMo, compared against ten different memory approaches including full-context, plain RAG, and other named systems.
- Read the actual numbers in the paper for latency and token cost reduction versus full-context baseline — this is the paper's central practical claim.
- Problem: existing memory systems (including Mem0-style approaches) use fixed, rigid operations and structures that don't adapt well across different task types.
- Mechanism: borrows from the Zettelkasten note-taking method (a system of small, atomic, densely cross-linked notes) — memories dynamically link to related memories, and the organization structure itself evolves as new memories are added, rather than following one fixed schema.
- Why this matters for your work: if your agent use case has facts that naturally interconnect in complex, evolving ways (rather than fitting a clean ADD/UPDATE/DELETE lifecycle), this paper's approach may be more relevant to you than Mem0's.
- Problem: the field was (and still is) fragmented, with different papers/libraries implementing incompatible memory models, making comparison and reuse hard.
- Mechanism: not a new memory algorithm itself — it's a unified library implementing many memory models (from various papers) under one consistent interface, organized in layers (basic functions → memory operations → full memory models).
- Why this matters: its comparison table is the single fastest way to see the whole landscape's tradeoffs in one place, rather than reading 10 separate papers.
- Read this as your "catch me up on 2026" paper — surveys covering mechanisms, evaluation methods, and open frontiers give you the vocabulary and map of the field as it currently stands, which is especially useful since this field moves fast enough that any single course (including this one) will be somewhat dated within months.
- LoCoMo: tests recall over long, multi-session conversations — the core "does the agent remember" test.
- LongMemEval: specifically designed around complex temporal reasoning (multi-hop time-based questions), closer to real enterprise usage patterns than simpler recall tests.
- DMR (Deep Memory Retrieval): the original benchmark introduced alongside MemGPT — an earlier, narrower precursor to the above.
- ConvoMem: conversational memory benchmark used prominently by Supermemory's self-reported results.
- For each of the 4 papers, write your own 3-sentence summary (problem / mechanism / result) without looking at the paper while writing — this forces you to actually retain the content, not just recognize it.
- Build a mini-benchmark: write 8-10 test conversations (in your chosen use case) each containing at least one fact that gets stated, then contradicted, later in the conversation. For each, write down the "expected" fact that should be retrievable at the end. This becomes your evaluation harness for Module 12.